CHANGELOG.md

Rama main · réplica de lectura del repositorio original.

# Changelog

All notable changes to this project are documented here.
Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).

---

## [Unreleased]

## [6.17.1] — 2026-11-08

**Issues fixed:** #244 (most questions got a `PersianMonthName = N'دی'` filter they never asked for); #237 (a data source variable missing from `.env` stopped the setup tools with no reason given).

### Fixed

- **A month, weekday or season name is found only as a whole word (`retrieval/value_retriever.py`) (PR #245).** Most questions got a `WHERE [d].[PersianMonthName] = N'دی'` filter they never asked for. The month, weekday and season names were looked for as a substring of the question with every space removed, so the two letters of «دی» were found inside ordinary words («جدیدترین», «تولیدی», «موجودی», «دیگر», «ردیف») and across word joins («تعداد یک» reads «تعدادیک» once the spaces go), «تیر» inside «تیرآهن», «آذر» inside «آذربایجان» and «مهر» inside «مهربان». A name now matches only when no letter is directly before or after it. Every spelling that named a real month, weekday or season still matches: spaces inside the name («پنج شنبه», «اردی بهشت»), ZWNJ («سه‌شنبه», «دی‌ماه»), «ماه» attached or not («دیماه», «دی ماه», «اسفندماه»), and digits right after it («اسفند۱۴۰۲»). Over the 2,346 Persian strings in the repository (example questions, golden cases, fixtures), the old and new matching differ on 39, and every one of the 39 was a false month filter that is now gone; no string lost a month, weekday or season it had matched before. The warehouse-vocabulary matcher was not affected: it ignores values shorter than three letters.
- **A data source variable missing from `.env` now says why (`core/dotenv_check.py`, `config.py`, `database/datasources.py`, `database/setup_sources.py`) (PR #241).** The setup tools (`python -m database.schema_inspector_cli` and the other tools that read `datasources.yaml`) stopped with only `environment variable Auction_DM_PASSWORD (named by password_env) is not set or is empty`, with nothing to show which `.env` was read or what was wrong with it. The message for a variable named by `username_env` or `password_env`, and for a `url_env` variable or `DB_CONNECTION_URL` that is unset or empty, now adds the reason, without printing any value (only paths, line numbers and variable names): which `.env` file was loaded, or that none was found, naming the folder the search really started from (the folder of `config.py`, or the working directory when python-dotenv searches from there: interactive, under a debugger or tracer, or frozen) and a `.env.txt` in it or above it as the likely cause, since a file manager that hides extensions saves that name; that `PYTHON_DOTENV_DISABLED` is set; a line python-dotenv could not parse; the same name written with different letters (`AUCTION_DM_PASSWORD` for `Auction_DM_PASSWORD`, where names are case-sensitive; on Windows they are not, so such a line is read as the variable itself); a name that is not in the file; a name present with an empty or blank value, or with a `${...}` reference that expands to an empty value when the variable it names is not set; a name that is in the file but not in the process environment (the process started before the file was edited, or another `.env` was loaded); an empty or blank variable in the process environment, which python-dotenv never overrides, reported first and also when there is no `.env` or the file does not assign it, with how to remove it (PowerShell and bash); and python-dotenv not being installed in the Python environment that runs the tool, in which case `.env` is not read at all and the message names that interpreter and says to install the requirements into it (`pip install -r requirements.lock`). `Settings.validate()` (so server start-up and the preflight) gives the same explanation for the same cases (a `url_env` value or `DB_CONNECTION_URL` that is set but wrong, such as a placeholder or the wrong dialect, is reported as before). The explanation never replaces the real error: if building it fails, the original message is raised unchanged. The errors for an unset variable are now `UnsetVariableError` (a `ValueError` subclass, same message). One more change outside the message: when python-dotenv is installed but `core/dotenv_check.py` cannot be imported, `.env` is still loaded (only the scan is skipped); before, nothing was loaded. No other setting is read differently.

### Documentation

- **The design documents describe 6.17 (`docs/design/DESIGN.md`, `docs/design/DESIGN-INVARIANTS.md`, `docs/design/RETRIEVAL.md`, `docs/admin-panel-architecture.md`) (PR #236).**
  - **Turn anatomy (`DESIGN.md` §5.2).** It gains the rating footer and the empty transcript's start questions. The resolved question is shown only when it differs from the question. The interpretation is listed in the main flow, where the code always put it (`DESIGN-INVARIANTS.md` §9 notes the correction).
  - **Decision log.** New entries D15 (both ratings on the answer), D16 (admin views), D17 (start questions) and D18 (Persian dates).
  - **`RETRIEVAL.md` §12.** It records the `ring_aliases` rule of 6.16.3.
  - **The admin architecture §3.2.** It describes the grouped views.
- **README (PR #236).** The feature table names the thirteen admin sections in their views, both ratings, the start questions and the filter-value fixes.
- **The GitHub Pages site (`site/`) (PR #236).**
  - **Three new cards in all five languages:** filter values spelled as the warehouse stores them, ratings and start questions, and the admin panel.
  - **The quick start** installs `requirements.lock` and ends with the preflight and `uvicorn` instead of the REPL.
  - **Phone widths.** The page no longer scrolls sideways there: the hero's glow reached 2 px past the edge.

### Upgrading

- **Restart the server; nothing to configure (PR #245, PR #241).** No setting is renamed or added, no file format changes, `requirements.lock` has no new pin (`git diff v6.17.0 -- requirements.lock` is empty), and the preflight has the same checks. The answer cache is in memory and starts empty on every restart, so no answer computed with a false month filter outlives the upgrade.
- **Some answers change, as they should (PR #245).** A question that named no month but contained «دی», «تیر», «آذر» or «مهر» inside another word («جدیدترین», «تولیدی», «تعداد یک», «تیرآهن») was filtered to that month; it now covers the whole period it asks for, so its figures are larger. A question that names a month («دی ماه ۱۴۰۳», «اسفندماه») is filtered exactly as before. One spelling is still read as a month: «تیر آهن» written with a space, where «تیر» is a separate word.
- **The setup tools explain a missing `.env` variable (PR #241).** If `python -m database.schema_inspector_cli` or the preflight stopped with `environment variable … is not set or is empty`, the message now adds why; it never prints a value.

## [6.17.0] — 2026-10-30

**Issues fixed:** none were open for these; they are usability work on both pages, asked for directly: the admin panel shows one view at a time with readable Persian dates, and an empty conversation offers the deployment's own example questions.

### Added

- **An empty conversation offers questions to start with (`GET /v2/examples`, `core/start_examples.py`, `web/js/main.js`) (PR #234).** Under «برای شروع، پرسش خود را در پایین بنویسید» the analyst page now shows up to six of the deployment's `examples.yaml` questions as chips («یا یکی از این‌ها را بپرسید:»); a click asks it. The new route returns question text only, never SQL or tags, and offers a question only when its SQL passes the guard with the caller's `denied_columns`: an example over a column the analyst may not read is not suggested to them, and one the guard refuses for any other reason (a write, an unknown table) is suggested to nobody. With no `examples.yaml`, no valid example, an older server or an error, the empty state is exactly as before. The demo mode's sample stories are unchanged.
- **The admin panel shows one view at a time (`web/admin/views.js`, `web/admin/main.js`, `web/admin/admin.css`) (PR #232).** Thirteen sections and the summary rail made one page about 6,800 px tall on a laptop screen. The sticky nav now opens with five buttons: «وضعیت» (maintenance, deployment checks), «کیفیت و استفاده» (audit, feedback, quality, usage), «دسترسی و کلیدها» (access requests, keys, failed authentication), «داده و پیکربندی» (cache, configuration, schema drift, vocabulary) and «همه», the page as it was. Only the chosen view's sections and jump links are shown; the summary rail is shown in every view. The panel opens on «وضعیت» and then on the view last chosen in that browser. A link to a section (`#sec-keys`, a bookmark) opens the view that holds it, and each summary-rail item now opens its section. A view that needs attention (a failed check, maintenance on, open feedback) has a dot on its button and the words «نیاز به توجه» for screen readers. Sections are hidden, not moved or unloaded, so every card loads and refreshes exactly as before; if setting the views up fails, every section stays shown.

### Changed

- **A turn no longer repeats the question as «برداشت سامانه از پرسش» (`web/js/render/turn.js`) (PR #234).** The box is shown when the resolved question differs from the question (a follow-up rewritten as a full question), and left out when it is the same text, spacing aside: it was the same sentence twice under the question bubble.
- **Admin dates are Persian calendar dates (`web/admin/time.js`, `web/admin/main.js`, `web/admin/access-requests.js`) (PR #232).** The audit range, the baseline date, the vocabulary card's `fetched_at` and an access request's creation time were printed as raw ISO strings (`2026-10-27T10:46:31.435668 → …`); they now read «از ۵ آبان ۱۴۰۵، ۱۰:۴۶ تا …», and the exact ISO string is the element's title (shown on hover). A string without an offset, as the audit log writes it, is shown on the browser's clock, which is the server's when the panel is opened where it runs. A value that is not a date is shown unchanged.
- **The deployment-checks card leads with what needs reading (`web/admin/main.js`, `web/admin/admin.css`) (PR #232).** FAIL and SKIP checks are cards, as before; the passed checks, usually most of them, fold into one line («۱۷ بررسی موفق») that opens to the same cards, so a PASS detail such as an `advice:` is one click away. The cache-age line no longer takes a cell of the card grid.
- **The summary rail writes «۲ از ۲۱» (`web/admin/main.js`) (PR #232).** «2 / 21» was laid out right to left in the Persian card and read as «21 / 2»; every rail figure now uses Persian digits, like the rest of the panel.

### Fixed

- **The admin top bar's backend-address field has its style back (`web/admin/admin.css`) (PR #232).** Its rules left `style.css` together with the analyst page's copy of the field, so the admin field and its «اتصال» button showed in the browser's default look; they now match the key controls beside them.
- **A race between two LLM warm-up tests (`tests/test_llm_warmup.py`) (PR #233).** The startup test released its blocked warm-up and ended without waiting for the thread, which still held the warm-up lock when the next test's warm-up started; on a slow runner that one sent nothing and failed. The module now waits for every warm-up thread a test started. Tests only; no product code changed.

### Upgrading

- **Reload the page after the upgrade (PR #232, PR #234).** The analyst page's and the admin panel's scripts and stylesheets are not versioned, so a browser that cached them keeps the old ones until a hard reload (Ctrl+F5).
- **Optional: make the start questions worth clicking (PR #234).** The empty conversation offers the `examples.yaml` questions whose SQL the guard accepts for that analyst. A question whose SQL names a table `schema.yaml` does not have (or a qualified column it does not list for that table) is offered to nobody, so the template's placeholder examples show nothing, and the empty state then looks as before. The questions are shown in file order, so put the ones a new analyst should see first at the top.
- **Nothing else changes.** No setting is renamed or added, no file format changes, `requirements.lock` has no new pin, and the preflight has the same number of checks. `GET /v2/examples` is a new read-only route behind the same key as every v2 route; the admin panel's sections, their ids and their refresh are as before, with «همه» showing the page as it was.

## [6.16.4] — 2026-10-29

**Issues fixed:** #223 (an «incorrect» button beside «پاسخ درست است»); #219 part 2 (after a clarification is answered, the page stays where it was while the new answer appears at the foot of the conversation). #219 part 1 (`Ring A`, `Ring B` offered as halls) is the template `session_policy.yaml`; `python scripts/draft_policies.py` (6.16.1) fills it from the real values.

### Added

- **«پاسخ نادرست است» on the answer, beside «پاسخ درست است» (`web/js/render/turn.js`, `web/js/render/feedback.js`, `web/styles/style.css`) (PR #230).** The wrong-answer flag («این عدد درست نیست؟») was only in the collapsed «جزئیات» drawer, so it was not found next to the confirmation. The new button in the result footer sends nothing itself: it opens the drawer and the flag form, puts focus on the category list and brings the form into view; the report is sent from that form, with the same categories, note and endpoint as before. Once sent, the button reads «✓ گزارش شد» (text and a critical border, not colour alone), and a later confirmation that replaces the report clears it, as it already reopened the form. It appears on every answer that has a result and a flag control (the confirmation's own rule is unchanged); on a narrow screen the two buttons wrap. The drawer's control is where it was.

### Fixed

- **After a clarification is answered (or an assumption edited) in live mode, the page moves to the new answer (`web/js/main.js`) (PR #230).** `PATCH .../assumptions` re-asks the question and the server answers with a new turn under its own `turn_id`; the page appended its card at the foot of the conversation but did not scroll, so the answer appeared out of sight. `rerenderTurn` now scrolls to the new card's result (to the card when it has none), as a streamed answer does once it settles. A turn that already has a card is still replaced in place, without a scroll. The page also keeps the record of the turn that was re-asked (its card stays on the page) and adds the new turn after it; before, the new turn was written over it.

### Upgrading

- **Reload the page after the upgrade (PR #230).** The web interface's script and stylesheet are not versioned, so a browser that cached them keeps the old ones until a hard reload (Ctrl+F5). Nothing else changes: no setting is renamed or added, no endpoint or file format changes, `requirements.lock` has no new pin, and the preflight has the same number of checks.

## [6.16.3] — 2026-10-29

**Issues fixed:** the root cause of #206 and #207 (a hall named in the question was filtered on a table name, `N'Ring'`, because the wizard had written table keys into `ring_aliases`); a deployment whose `aliases.yaml` came from the wizard must also restore its ring names (see Upgrading).

`ring_aliases` in `aliases.yaml` maps the **exact stored value** of the ring dimension's name column to the forms analysts type for it; `ValueRetriever.extract_ring` finds a form in the question and the key becomes the ring filter (the session's default scope, `WHERE r.[Name] = N'<key>'`). The setup wizard of 6.13 to 6.16.2 wrote one entry per **table** there instead (key = the table's key, values = the table's aliases): on one deployment 75 entries keyed by table. A question that named a currency filtered the ring by `N'Currency'`, one that asked for the halls by `N'Ring'`, a trade question by `N'Schema.Trade'`, and the answer was empty; the preflight passed that file (`Aliases are curated: 0 synonym key(s), 75 ring alias(es)`). The same deployment's earlier curated file (real ring names with their forms) gave the right exact names for the same questions. A file whose keys are real ring names behaves exactly as before (pinned by `tests/test_ring_aliases_pins.py`, written before the change and unchanged since).

### Fixed

- **A `ring_aliases` entry keyed by a table is ignored when `aliases.yaml` loads (`core/ring_names.py`, `knowledge/aliases.py`) (PR #228).** An entry whose key is, as a whole and with letter case ignored, a key of `schema.yaml`'s `tables` is left out of `RING_ALIASES`, and one `WARNING` names how many and which (at most ten, then `and N more`); no alias is logged. Only that exact match counts: a ring called `North Hall` is not touched because a table is called `Hall`, and a bare name that merely resolves to a table key (`Customer` for `sales.Customer`) is not a match either. When `schema.yaml` is absent or does not parse, or the guard fails for any reason, every entry is used as written, as before; when no key is a table key the very same mapping is returned.
- **The setup wizard no longer writes a table's aliases into `ring_aliases` (`setup_project.py`, `core/wizard_drafts.py`) (PR #228).** Step 3 keeps a table's aliases in `entities.yaml`, where it always also wrote them and where retrieval reads them, and leaves `ring_aliases` as it found it (`{}` on a first run). When it merges into an earlier `aliases.yaml` it drops the entries whose key names a table (a `schema.yaml` key, a bare name that finds a key, or a `Schema.Table` spelling of a table keyed differently) of this run, of the earlier `entities.yaml` or of the output folder's `schema.yaml`; the aliases of a dropped entry are added to that table's entity (an entity is made for a table that had none), and the summary and `.setup_log.json` (`step6_write.dropped_table_rings`: `count`, `keys`, `aliases_added_to_entities`, `entities_created`, `unplaced`) give the count and the keys, which are table names. Every other entry, a ring name, is not touched, whatever words it shares with a table. A `--resume` no longer reuses the table-keyed copy a log of 6.16.2 holds.
- **The preflight check `Aliases are curated` fails on a `ring_aliases` key that is a `schema.yaml` table key (`scripts/verify_deployment.py`) (PR #228).** `N of M ring_aliases key(s) are keys of schema.yaml tables: ...` (up to ten keys, then `and N more`), the exact entries the application ignores, with the way out. The check count is unchanged. When the running application has cached the ring dimension's values (the admin panel's card; the command line has no cache) the `[PASS]` detail adds `advice:` with the number of keys that are not among them, as a count only. The advice of `Entities match schema keys` no longer tells you to give `ring_aliases` the same keys as the entities.

### Added

- **`scripts/draft_policies.py --ring-aliases` drafts `ring_aliases` from the stored ring values (`scripts/draft_policies.py`, `core/policy_drafts.py`) (PR #228).** It adds, to a copy of `aliases.yaml` written to `project_config_draft/`, an entry for each stored value of the scope column that has none: the key is the exact value, the forms are the value itself and, with `--generic-word WORD` (repeatable), the value without that leading word when a whole word is left and no other value gives the same form. New entries are written longest value first (the first entry with a form in the question wins). Existing entries, `synonyms`, comments and line endings are kept byte for byte, the result is loaded with the application's validator and compared with the input before anything is written, and an inline `ring_aliases: {...}` is refused, not rewritten. It prints counts, including existing keys that are table keys (left as they are) and existing keys that are not among the stored values. Nothing changes without the option.

### Changed

- **The wizard's `ring_aliases` is empty (PR #228).** The six assertions in `tests/test_setup_wizard_output.py` and `tests/test_setup_wizard_drafts.py` that expected a table-keyed `ring_aliases` from a run now expect none.

### Upgrading

- **If your `aliases.yaml` was written by the wizard of 6.13 to 6.16.2, do this (PR #228).** On start the log has `aliases.yaml: N ring_aliases key(s) are keys of schema.yaml tables and are ignored` and `python scripts/verify_deployment.py` fails `Aliases are curated` with the same count. The application already ignores those entries, but a hall you want recognised needs its own entry. Either restore the curated ring names from a backup, or run `python setup_project.py` again over the folder that holds the file (`--output`; with `--overwrite-live` for the live folder, which backs it up first): it drops the table-keyed entries, keeps their aliases in `entities.yaml` and reports the count and the keys. Then add the real ring names, by hand or with `python scripts/draft_policies.py --ring-aliases [--generic-word WORD]` (read the forms, then copy the file across), and run the preflight again.
- **A file whose `ring_aliases` keys are real ring names is not changed at run time, and the preflight still passes it (PR #228).** The one case that differs: a ring whose stored name is exactly, letter case aside, a key of `schema.yaml`. The rule cannot tell it from the wizard's entry, so it is ignored and the preflight fails. There is no setting that switches the guard off; if you have such a ring, report it on the issue tracker.
- **Nothing else changes (PR #228).** No setting is renamed, no file format changes, `requirements.lock` has no new pin, and the preflight has the same number of checks.

## [6.16.2] — 2026-10-29

**Issues fixed:** #214 (an attached plural such as «تالارها» was not recognised), #218 (a filter value written with Persian ی/ک found nothing when the column stores Arabic ي/ك).

Two reports of the same kind: the question and the model write Persian text, and the configuration and the warehouse hold it in a slightly different spelling. A statement filtered on a hall name returned nothing because the warehouse stores `ي`/`ك` (Arabic) where the model writes `ی`/`ک` (Persian), and a question that wrote a plural attached to its word (`تالارها`) did not find what `تالار` or `تالار ها` found. Neither change removes anything a request does today: the first acts only on a literal the column does not contain, the second only adds a term (pinned by `tests/test_filter_spelling_pins.py` and `tests/test_plural_matching_pins.py`, written before the change and unchanged since).

### Fixed

- **A filter literal written with the Persian letters now finds a value the warehouse stores with the Arabic ones (`retrieval/filter_spelling.py`, `retrieval/filter_values.py`, `core/persian.py`, `llm/sql_agent.py`, `session/engine.py`, `config.py`) (PR #226).** `WHERE r.[Name] = N'<value with ی>'` returned no rows while the same statement with the value as stored (`ي` where the model wrote `ی`) returned them, because under the deployment's collation the two strings are different. The filter value check of 6.16.0 could not help: it compares normalised forms, so it counts the Persian spelling as a known value and changes nothing. A statement the guard has accepted is now read once more before it runs, with no model call: for `=`, `<>`/`!=`, `IN` and `NOT IN` on a column whose whole value set is cached (a table's `prefetchable_columns`, the rules of the filter value check: a missing, empty or possibly truncated list, a denied or join-only column and a column that is not a plain read of a stored one are never touched), a string literal that is **not byte-equal to any stored value** but equals **exactly one** of them up to letter form is replaced by that stored value, character for character; two or more candidates, none, or another stored value that only differs from the literal by what the question matching folds (a trailing space, case) leave it as written. "Up to letter form" is `core.persian.letter_form_key`: `ي` and `ى` with `ی`, `ك` with `ک`, Arabic presentation forms and ligatures to their base letters, canonical composition (NFC), and ZWNJ with none (`می‌خواهم` and `میخواهم`). Nothing else is folded: not case, not digits (`۱۴۰۲`, `١٤٠٢` and `1402` stay three different literals), not spaces (a ZWNJ is not a space), not `ة`/`ه` (different letters), `ۀ`/`هٔ` (a letter and an optional mark) or the alefs. For `LIKE` and `NOT LIKE` on such a column the variant letters of the pattern are written as the column stores them when, for that group (`ی ي ى` or `ک ك`), every stored value that has one of its letters uses the same form; a column that stores both forms of a group, or an Arabic presentation form anywhere, is left alone (an `OR` expansion would change the shape of the statement the model wrote; that case is left to the model and the filter value check). The edit replaces only the characters of the literal's own token (`N` prefix and quoting kept, the rest of the statement byte for byte), the result is parsed again and must read as the original with just those literals changed, and the guard validates the rewritten statement again in the dialect it runs in; if either refuses, the original runs. It is used in `/query` (mode `result` and `full`) and in a session turn's fresh and carry-forward generation, on every statement that is about to run, including the corrected statement of the filter value check; mode `sql` (nothing runs) and the `_prev` refinement are not touched. With `FILTER_VALUE_CHECK` on it runs first, so a spelling difference is a known value and costs no correction round. `result.sql`, the SQL the UI shows and the audit record's `generated_sql` are the statement that ran (`SQLGenerationResult.raw_response` is still the model's text).
- **A Persian word written with an attached plural is matched as its stem too (`core/persian.py`, `schema_data/retriever.py`, `retrieval/table_evidence.py`, `retrieval/source_selector.py`) (PR #226).** `تالارها`, `تالار‌های` (with the half-space, which the normaliser folds to the attached form) and `تالارهایی` were one token that no description, trigger phrase, synonym or data-source keyword contained, while `تالار ها` (with a space) put `تالار` in a token of its own and matched. `core.persian.strip_plural_suffix` gives the stem of a word with one of the endings `ها`, `های`, `هایی`, or `ها` with a possessive (`هایم`, `هایت`, `هایش`, `هایمان`, `هایتان`, `هایشان`), when what is left is at least three letters; nothing in Latin script; no `ان` or `ات`, which end `تهران`, `ایران` and `حیات` as well as plurals. It is used **in addition to** the word, never instead of it: the TF-IDF scorer adds the stemmed question's terms and bigrams (`rank_tables`), the `always_include` triggers also try the stemmed question (`forced_tables`), a synonym is found under the stem as well as under the word, the column evidence tries the stems (so `هایی` and the possessives are known there too, where only `ها` and `های` were), an alias of several words in `entities.yaml` or `fact_patterns` is also compared with such words reduced (`alias_hits`), and the data-source `keywords:` are also looked for in the question with such words reduced (`keyword_hits`, still whole words: `تالار` matches `تالارها`, not `تالاری`). A question without such a word is matched and scored exactly as before; for one with such a word every table that scored still scores, and no lower, though a table that now scores can rank above it and, with more than six candidates, push it out of the top six. The matching of a stored value against the question (`retrieval/dimension_vocabulary.py`) is not changed: it does not use this tokenizer, it already finds a value inside a longer word, and an extra term there could make two values tie.

### Added

- **`FILTER_VALUE_SPELLING` (default `true`) switches the spelling fix off (`config.py`, `.env.example`) (PR #226).** It is on by default because it acts only where the statement would otherwise compare with a value the column does not contain, and what it writes is the one stored value the literal is a spelling variant of; it issues no query and starts no background refresh (the vocabulary cache is only read), and a column without a cached vocabulary is left as it is. Set `FILTER_VALUE_SPELLING=false` and every statement runs exactly as the model wrote it, every audit line, prompt and response is as in 6.16.1 (pinned).
- **`value_spelling` in the audit record (`observability/audit.py`) (PR #226).** Written only when the fix rewrote a literal: `{"fixed": 1, "columns": ["Ring.Name"]}`, how many literals and the columns they are compared to. Neither the literal nor the stored value it became is written, the same names-never-values rule as `value_check`; a block with any other key or type raises. With the setting off, or nothing rewritten, the key is absent and the line is byte for byte what it was; lines written before this release load as before.
- **`core.persian.letter_form_key`, `core.persian.strip_plural_suffix` and `strip_plural_suffixes`, and the shared helpers `usable_vocabularies`, `comparison_literals` and `like_patterns` of `retrieval/filter_values.py` (PR #226).** `normalize_for_matching`, `normalize_compact` and `NORMALIZER_VERSION` are unchanged, so the query cache keys are too (pinned). The filter value check now uses `usable_vocabularies` for the rule it had inline; its results are unchanged.

### Upgrading

- **One behaviour is on by default (`FILTER_VALUE_SPELLING=true`) (PR #226).** No setting is renamed, no file format changes and `requirements.lock` has no new pin. A literal that is a stored value, a column with no cached vocabulary, a literal two stored values could be, and a `LIKE` over a column that stores both forms all run as written. If you would rather see every statement exactly as the model wrote it, set `FILTER_VALUE_SPELLING=false` and restart.
- **Reading the audit log (PR #226).** `value_spelling` present means the statement that ran (`generated_sql`) has literals the model wrote differently; the model's own literal is not kept anywhere in the record. Requests it fixed spend no correction round, so they do not raise the "needed a correction" share. The cache it reads is the one `FILTER_VALUE_CHECK` reads: a list that is stale is used as is, so a value added to a dimension since its last refresh is not known to the fix, and a literal is only ever rewritten to a value the list holds.
- **Retrieval for a question with an attached plural can change (PR #226).** A table whose description, trigger phrase or synonym holds the singular is now found for `...ها`, `...های`, `...هایی`; tables that were found are still found. Check a question you rely on with the plural written attached.

## [6.16.1] — 2026-10-29

**Issues fixed:** the first part of #219 (the template's «Ring A» / «Ring B» options in the default scope and the analysts' memory).

An operator reported that "the file for the analysts' memory was removed entirely". Nothing was removed: `project_config/memory_policy.yaml` (the closed set of values an analyst may pin) and its sibling `session_policy.yaml` (the scope a ranking question gets when it names none, and the options its clarification offers) are two of the ten required files, but no setup tool wrote them, and the preflight only checked that they load. On a first install both were the template, whose options are made-up samples (`Ring A`, `Ring B`, `All rings`), so every real value an analyst tried to pin was outside the closed set and refused, and the clarification offered values the warehouse does not have. The feature looked gone. This release writes the real values and says when they are missing. No runtime behaviour of memory or of the default scope changes.

### Added

- **`scripts/draft_policies.py`: draft the real values of `memory_policy.yaml` and `session_policy.yaml` from one dimension column (`scripts/draft_policies.py`, `core/policy_drafts.py`) (PR #224).** `--table` and `--column` pick the dimension, by default the `filter_key` of `session_policy.yaml` and the `column` of the memory key that shares it when they name a real table and column of `schema.yaml` (a table is found as the application finds a configured table name, bare or `Schema.Table`). It reads the column's distinct values with one bounded statement (`SELECT DISTINCT TOP (N+1) ... ORDER BY`, `--max-values`, default 200; a column with more is refused as not a small dimension) through the application's own engine and login, on the data source of the table's `datasource:`. It keeps the values the wizard's rules accept (`core.value_shapes`: no date, number, code, phone number, address, masked value or error text; a column named like personal data is never read), without the session policy's `default_label`, one spelling of each value under the project's Persian and Arabic folding, sorted. It writes `project_config_draft/memory_policy.yaml` and `project_config_draft/session_policy.yaml`: the current files with only `options` replaced (and `filter_key` / `column` when the dimension is moved); comments, `max_length`, `field_name`, `default_label`, `clarification_prompt` and every other field stay byte for byte, the memory options never contain the `default_label`, and the result is loaded with the application's validators and compared with the input before anything is written. Nothing is written into `project_config/` or `project_config.example/` without `--overwrite-live` (then each replaced file is first copied to `<name>.bak-<stamp>`); an existing draft needs `--force`. It prints counts only; `--show-values` adds the first ten values, `--dry-run` writes nothing. It is a separate command, like `draft_descriptions.py`, because it reads one column and asks no model, and has to be run again when the dimension gains a value.
- **A preflight check, `Session and memory policies use real values` (`scripts/verify_deployment.py`, `core/policy_values.py`, `core/config_doctor.py`) (PR #224).** It fails, naming the file and the key and never a value, when a `filter_key` names no table of `schema.yaml`, when a memory key's `column` is not a column of that table, or when the options of a file are exactly the sample values of `project_config.example/` (compared as a set), and points to the new command. The shipped example is not compared with itself, so it passes its own preflight. The PASS detail advises, as counts, how many options are not among the dimension values the running application has cached (the admin panel's card; the command line has no cache). File reads only, so the admin panel runs it without "deep checks". The preflight now has twenty-six checks with one data source (twenty-seven with `FILTER_VALUE_CHECK`).
- **Documentation (PR #224).** Runbook section 2.5 and the check's row in section 3.1; section 0.3.4.1 and the check's row in both setup guides; the first-install checklists name the command in the wizard's step; `project_config.example/README.md` and the README list the script.

### Upgrading

- **Nothing changes at run time (PR #224).** No setting is renamed, no file format changes, `requirements.lock` has no new pin, and the server answers exactly as before with the files you have.
- **The preflight can now fail on a configuration that works (PR #224).** If `project_config/` still holds the template's `memory_policy.yaml` or `session_policy.yaml` (a first install, or an upgrade that never replaced them), the new check fails, and so does a `filter_key` that is no table of `schema.yaml`, though the server runs. Run `python scripts/draft_policies.py`, read `project_config_draft/memory_policy.yaml` and `project_config_draft/session_policy.yaml`, copy them into `project_config/` and run the preflight again. Files that already hold real values and a real table pass untouched.
- **`filter_key` stays the key the retriever stores the scope under (PR #224).** A ring named in a question reaches the session under the key `Ring` (`retrieval/value_retriever.py`), so `--table` should be used to move `filter_key` only when you know your retrieval stores the scope under that key; without `--table` the command leaves `filter_key` as it is and only refreshes the options.

## [6.16.0] — 2026-10-27

**Issues addressed:** #204, #206, #207 (filter values the model invented for the ring table); the check is opt-in, and the deployment's configuration is part of the fix (see Upgrading).

The model sometimes invents the value it compares a dimension's label column to: `WHERE r.[Name] = N'Ring'` (the table's own name), a phrase of the question that names no ring (`N'online channel'`, which is a whole source database, not a ring). The statement is valid, the guard has no reason to refuse it, and the answer is empty or wrong. This release adds an opt-in check for it. It is **off by default**; with it off no request, prompt, statement, audit line or response differs from 6.15.2 (pinned by `tests/test_filter_value_check_pins.py`).

### Added

- **`FILTER_VALUE_CHECK` (default `false`): a filter value that does not exist gets one correction round instead of an empty answer (`retrieval/filter_values.py`, `llm/sql_agent.py`, `session/engine.py`, `config.py`) (PR #212).** With it on, a statement the guard has accepted is read once more before it runs: the string literals it compares to a column by `=`, `<>`/`!=`, `IN` or `NOT IN` are resolved to the stored column they read (through table aliases, CTEs and derived tables) and compared with the value set `retrieval/dimension_vocabulary.py` already caches for a table's `prefetchable_columns`, with the same normalisation as a question is matched with (Persian and Arabic letter forms, digits, ZWNJ, case, whitespace). A literal that is not among the values gets ONE correction message through the usual correction layout (appended to the question segment; the static prefix is untouched, so the prefix cache is not disturbed): the column, that the value does not exist, the nearest real values (up to five, character trigrams, deterministic; fenced as warehouse data like every value the prompt quotes), and "if the question does not name a specific value of this column, remove this filter". The corrected statement runs only if it passes the guard, keeps no unknown literal (and, in a session turn, no filter the question named) and executes. Anything else (the model keeps the value, the guard refuses the new statement, the model or the database fails) runs the original statement exactly as it would have run without the check, so the check can add one model call but never refuses a request. It counts in the correction counter (`llm.corrections`, `correction_attempts`). It is used in `/query` (mode `result` and `full`) and in a session turn's fresh and carry-forward generation; mode `sql` (no correction loop) and the `_prev` refinement are not checked. It reads the cache only: no database read on the request path, a column with no cached values is not checked (a stale list is used, and the existing background refresh is started for a missing or stale one), and a list as long as the prefetch cap (`DEFAULT_TOP_N` distinct values) is not trusted to be complete. A column the key's `denied_columns` denies, or restricts to join-only, is never checked, never named in the message and none of its values is offered.
- **`security.sql_guard.column_origins`: the stored columns a column reference reads (`security/sql_guard.py`) (PR #212).** The resolution the guard does for a denied or join-only column, with the table named, through aliases, derived tables, CTEs, `*` and set operations. Nothing in `validate_sql` calls it, so no allow or deny decision changes.
- **`value_check` in the audit record (`observability/audit.py`) (PR #212).** Written only when the check ran and found something to compare: `{"checked": n, "unknown": [{"column": "Ring.Name", "value_sha256": "<16 hex>", "length": 10}], "corrected": bool, "fallback": bool}`. The literal is not written (this log keeps names, never values; the statement that ran is in `generated_sql` as ever) and neither are the real values offered to the model; the hash lets the same invented value be counted across requests. `fallback` is true when there were unknown literals and the original statement ran. The record cannot be given a value to write: a block with any other key or type raises, as `columns` and `resolved_columns` do. With the check off, or with nothing to compare, the key is absent and the line is byte for byte what it was; a reader treats a missing key as no check.
- **Preflight advice `Filter value check has vocabularies` (`scripts/verify_deployment.py`, `core/config_doctor.py`) (PR #212).** Listed only while `FILTER_VALUE_CHECK` is on, so the preflight is unchanged with it off. It names the dimension tables (those `entities.yaml` names) that have label-like columns (`Name`, `Title`, `Kind`, `Status`, `Type` ..., not a person's name) and no `prefetchable_columns`: the check cannot help there. Advice only (a `[PASS]` whose detail starts `advice:`, table and column names only), never a `[FAIL]`; file reads only, so the admin panel runs it too.

### Upgrading

- **Nothing changes until you turn it on (PR #212).** No setting is renamed, no file format changes, `requirements.lock` has no new pin, and `FILTER_VALUE_CHECK` defaults to `false`. To try it, set `FILTER_VALUE_CHECK=true` in `.env` and restart.
- **It needs `prefetchable_columns` (PR #212).** The check compares with the value sets the dimension vocabulary caches, and those exist only for the columns a table lists under `prefetchable_columns` in `schema.yaml` (small dimensions only; leave out tables of people and anything large). Run the preflight with the setting on: the advice line `Filter value check has vocabularies` lists the dimension tables that have label-like columns and no `prefetchable_columns`. With `DIMENSION_VOCABULARY_WARM_ON_STARTUP=true` (the default) the sets are cached at start-up.
- **What it costs (PR #212).** At most one more model call per request or turn, only when a literal is not a known value and `MAX_CORRECTION_ATTEMPTS` is not zero (outside that budget: a request that used all of it can still get this one call). It is not a refusal path: when in doubt the original statement runs. A value that is real but not yet cached (a ring added since the last refresh) costs the extra call and then runs as written.
- **Reading the audit log (PR #212).** `value_check.unknown` non-empty with `corrected: true` is an invented value that was replaced; with `fallback: true` it is one the model kept (or whose correction failed) and that ran as written. Requests it corrected count in `llm.corrections`, so the real-use quality card's "needed a correction" share rises by them. Lines written before this release, and lines written with the check off, have no such key, and every reader treats it as absent. The quality card has no figure of its own for the check.

## [6.15.2] — 2026-10-27

**Issues fixed:** #183 (schemas and tables to leave out of discovery), #203 (a ring id drawn as a chart), #205 (volume reported in Rial).

A patch for three things seen on a deployment: the answer summary put Rial/ریال on volumes and counts, a list of ids was drawn as a chart, and the setup tools could not skip a single table or keep a staging schema's tables without its change-capture tables. Every existing file behaves as before; the new keys are optional.

### Fixed

- **The answer summary no longer writes Rial/ریال on numbers that are not money (`llm/interpret.py`, `llm/column_units.py`, `api/runner.py`) (PR #210).** The summary prompt said that every monetary value is in Rials and that amounts are always written with the unit Rial/ریال, and the model put the unit on every number it saw: a total traded volume of `33,493,503,232` came back as "33,493,503,232 Rial", and «حجم کلی معاملات چه قدره؟» as "۸۲,۱۹۰,۶۶۳ ریال". The rule is now "Rial/ریال only for a value that is money, never for a count, quantity, volume, weight, percentage, id, code or date", and the prompt lists, for each numeric column of the rows shown, whether it is money: first from the column's `column_units:` in `schema.yaml` (below), else from its name (`price`, `value`, `amount`, `cost`, `fee`, `ارزش`, `قیمت`, `مبلغ`, `کارمزد` and the like are money; `count`, `quantity`, `volume`, `weight`, `percent`, `id`, `code`, `date`, `حجم`, `تعداد`, `درصد`, `کد` and the like are not; a name with both, such as `FeeCount`, is not), else "not known" and the model decides. A money column is still asked for in Rial/ریال, the replacement of toman/تومان by Rial/ریال in the answer is unchanged, and so are the thousands separators and the 20-row preview. `/query` (`interpret: true`) and the conversational engine share one prompt; the unused copy of the template in `api/runner.py` is now a name for the shared one. Tests: `tests/test_interpret_prompt_pins.py` (what must not change), `tests/test_interpret_units.py`.

- **A numeric identifier column is no longer drawn as a chart measure (`web/js/num.js`, `web/js/render/table.js`, `web/js/render/chart.js`, `web/js/chart-engine/recommend.js`) (PR #210).** "All trading rings with their codes" returned (Name, ID) rows, `determineShape` took the one numeric column of the two as the measure, and the ids were drawn as a bar chart. A numeric column whose name is an identifier is now a label: such a result is a table. The rule, `isIdentifierName` in `web/js/num.js`, reads the name only: `ID`, `Id` or `id`; `id` or `code` as the last word after a space, `_`, `-` or `.` (`ring_id`, `Broker Code`); an upper-case `ID` after a lower-case letter or a digit (`RingID`); a capitalised `Id` or `Code` after a letter or a digit (`ringId`, `BrokerCode`, `ISOCode`); and the Persian word `کد` or `شناسه`. A name that merely ends in the same letters is not an identifier (`Paid`, `Valid`, `Width`, `Barcode`, `Postcode`), and neither are all-capital names such as `PAID`, `VALID` and `BROKERID` (write `BROKER_ID`). Only a numeric column is affected, so a text `Code` still labels a chart. Every other shape decision is unchanged: a one-row result is still a record, a 1 × 1 numeric result is still a scalar (also when the column is an id), two numeric columns are still a table, and the 2 to 30 row range is the same. `chart.js` takes the first numeric non-identifier column as the measure, and `chart-engine/recommend.js`, which may not import and keeps its own copy of the rule, recommends nothing for a result whose only numeric column is an identifier. No server code decides a result's shape, so nothing changes in the API. Documentation: `docs/design/CHART-ENGINE-BOUNDARY.md`. Tests: `tests/web_ui/test_web_ui_shape_pins.py` (what must not change), `tests/web_ui/test_web_ui_identifier_columns.py`, scenario 6 of `tests/web_ui/run_result_shapes.mjs`.

### Added

- **`exclude_tables:` per source and `EXCLUDE_TABLES` for every source: skip single tables when the setup tools discover tables (`database/schema_filter.py`, `database/datasources.py`, `config.py`, `database/schema_draft.py`, `schema_data/sync.py`, `setup_project.py`, `scripts/sync_schema.py`, `scripts/verify_deployment.py`) (PR #210).** `exclude_schemas:` skips a whole schema; a deployment that wanted to skip `X_Fact.Order`, or the `_Delete` tables of one schema, had no way to say so. The patterns are `fnmatch`, compared without regard to case with `schema.table` (`X_Fact.Order`, `X_Stg.*_Delete`, `*.Order`); a pattern with no `.` is compared with the table name alone (`Order`, `*_Delete`). `EXCLUDE_TABLES` (comma-separated, `.env`) applies to every source and the source's `exclude_tables:` list to that source, exactly as `EXCLUDE_SCHEMAS` and `exclude_schemas:` do. They are used where those are: the first-install inspector (`database.schema_inspector_cli`, new rule `configured-tables` in its summary), the wizard (counted as `named by a table pattern`, also in `.setup_log.json` as `excluded_tables`), and `sync_schema.py` (a table they name is neither listed nor added by `--add-tables`; the report has the section `tables named by an excluded table pattern, not listed or added`). Discovery only, like `exclude_schemas:`: a table `schema.yaml` already lists is synced as before, and the preflight check `No table in an excluded schema` now also lists a listed table that a pattern names (its text is unchanged when only schemas are involved). The legacy `SchemaInspector.inspect` that `AUTO_DISCOVER_SCHEMA` uses at run time does not read `EXCLUDE_TABLES`. `database/datasources_edit.py` gains `add_exclude_tables`, `record_exclude_tables` and `exclude_tables_snippet` (the same comment-preserving, backed-up edit as for `exclude_schemas:`); the wizard does not ask for table exclusions to record. Documentation: `.env.example`, `project_config.example/datasources.example.yaml`, runbook §2.3 and §16.3, section 0.3.3 of both setup guides, `docs/design/DATASOURCES.md`. Tests: `tests/test_exclusion_pins.py` (what must not change), `tests/test_exclude_tables.py`.
- **`keep_staging_schemas:` per source: scan a staging schema without its change-capture tables (`database/schema_filter.py`, `core/config_doctor.py`) (PR #210).** `STAGING_PATTERNS` drops a staging schema (`*_Stg`, `*stg*`, `*staging*`) and the change-capture tables (`table:*_Inserts`, `table:*_Updates`, `table:*_Deletes`) together, and `--include-staging` keeps both. A schema named in the source's `keep_staging_schemas:` is exempt from the schema patterns, so its tables are drafted, while the `table:` patterns and `exclude_tables:` still remove its change-capture tables; `exclude_schemas:` still wins. The default `STAGING_PATTERNS` are unchanged and still know the plural names only; the singular `_Insert`, `_Update`, `_Delete` are dropped with `exclude_tables: ["X_Stg.*_Insert", ...]` or by adding `table:*_Insert,table:*_Update,table:*_Delete` to `STAGING_PATTERNS`, as the runbook now says. The preflight check `No staging or change-capture tables` does not count a kept schema's tables against the schema patterns. `core.config_doctor.is_staging` and `staging_tables` take the keep list as an optional argument.
- **`column_units:`, an optional per-table map `{column: rial | none}` in `schema.yaml` (`schema_data/registry.py`, `schema_data/sync.py`) (PR #210).** It says whether a number in a column is money, for the wording of the answer summary only: the prompt's schema block, the SQL guard's allowlist and every query are unchanged. `columns:` keeps its `{name: description}` shape, which every reader and the sync's line editor rely on, so the unit is a map of its own beside `column_types:`; a file without it loads and behaves as before, and a file with it still loads on an earlier release, which ignores the key. Values are `rial` or `none` (any letter case; write `none`, not `no`, which YAML reads as false), keys must be columns of the table's `columns:`, and a result column is looked up by its name in the result (a name that two tables give different units is not looked up). The sync leaves the map as it is; `sync_schema.py --prune` removes the entry of a column it removes, so the file stays valid. `schema_data.registry` gains `get_column_units` and `COLUMN_UNITS`. Documentation: the header and the `Order` table of `project_config.example/schema.yaml`, `project_config.example/README.md`, runbook §16.3, section 0.3.3 of both setup guides. Tests: `tests/test_schema_column_units.py`, `tests/test_schema_sync_units.py`.

### Upgrading

- **No setting is renamed, no file format changes and `requirements.lock` has no new pin (PR #210).** The three new keys, `column_units:` in `schema.yaml` and `exclude_tables:` and `keep_staging_schemas:` in `datasources.yaml`, are optional and change nothing when absent; `EXCLUDE_TABLES` is a new optional setting. The SQL the server writes, the guard and the audit log are as in 6.15.1.
- **Where the summary's money guess is wrong, say it in `schema.yaml` (PR #210).** Write `column_units: {Quantity: none, TotalPrice: rial}` under the table. A column that is not listed is judged by its name, as described above.
- **To skip single tables or keep a staging schema's tables, edit `datasources.yaml` and re-run the setup tool (PR #210).** The default `STAGING_PATTERNS` is unchanged, so earlier runs propose the same tables; singular change-capture names (`*_Insert`) need an `exclude_tables:` entry.

## [6.15.1] — 2026-10-27

**Issues fixed:** #202 (a `denied_columns` entry that did not hide the column it named).

A security patch for column restrictions. On a deployment, a key that denied `NationalID` still saw it in web answers. The guard refuses that column when the entry is written exactly, but an entry with a space or an invisible character at either end (easy to paste from a Persian editor) denied nothing, a plain name that matches no column was never reported, and the audit log of the web UI did not say which key asked. Entries are now cleaned at their ends, unmatched names fail a new preflight check, look-alike columns are reported as advice, and every web turn records `principal_id`. Every key that works today still works.

### Security

- **A `denied_columns` entry with a space or an invisible character now restricts the column it names (`security/column_policy.py`, `security/auth.py`, `appdb/key_store.py`) (PR #208).** The guard rejects a statement that selects a denied column, but an entry such as `"NationalID "`, `" NationalID"`, `"NationalID<RLM>"`, `"<ZWNJ>NationalID"` or one wrapped in a bidi embedding or isolate or a byte-order mark never equalled the column name, so it denied nothing, and nothing logged that. Persian editors insert U+200C, U+200E, U+200F, U+202A to U+202E, U+2066 to U+2069 and U+FEFF without showing them, so a list copied from a document could be fail-open and look right. Every entry is now normalised in one place, `normalize_denied_entry`, before it is read: whitespace and Unicode format characters (category `Cf`) are removed from the **start and the end** of the entry, repeatedly, so a run such as a mark, a space and a non-joiner goes whole. Nothing inside an entry is changed, format characters included: a Persian column name may itself hold a zero-width non-joiner, and the entry stays exactly what the operator wrote in the key file, in the panel, in the stored list and in the `PATCH` response. Letters, case and inner spaces are not touched either (`Price List` stays `Price List`, and an entry with an inner space that no column has is reported, below). For a plain entry that does hold a format character inside (`National<ZWJ>ID`), the policy the guard matches against denies the entry as written and also the same name without that character, so the restriction can only be wider than the entry, never narrower; the second name is not written back anywhere. The file and `API_KEYS_JSON` keys, the keys of the admin panel (when they are read and when `PATCH /admin/keys/{key}/acl` writes them), the access-request approval and the policy parser all go through it, and so does `Principal`, so a principal built any other way is cleaned the same. A scoped entry (`Table.Column`, `Source:Table.Column`) is cleaned at its two ends the same way before it is checked against `schema.yaml`.
- **An entry that normalisation changed is logged once per key, as a WARNING that names the key id and shows the entry with `ascii()` (PR #208)** (so the invisible character appears as `\u200f`). The key's hash is never logged.

### Fixed

- **A plain `denied_columns` name that matches no column of `schema.yaml` is reported instead of silently denying nothing (`security/auth.py`, `scripts/verify_deployment.py`) (PR #208).** At load, one WARNING per key per process lists the entries that match no column (the match ignores case, as the guard does; an entry matches when it, or it without the format characters inside it, is a column name). The server still starts, so no installation that starts today is stopped. The new preflight check below finds them again, with the key id.
- **The audit records of the web UI carry `principal_id` (`session/engine.py`, `api/v2_routes.py`) (PR #208).** Every record written for a turn of `/v2/sessions` had `principal_id: null`, because the turn engine built the record without it (the `/query` path always set it), so the audit log could not say which key asked what. A first turn, a follow-up, a streamed turn and the re-run of `PATCH .../assumptions` now write the caller's id. `TurnEngine.ask` takes an optional `principal_id`; without it the field stays empty, as before. Existing fields are unchanged and earlier records load as they did. Records written by 6.15.0 and before keep the `null`; the other audit writers are listed under Upgrading.
- **Approving an access request removes the entry whichever way it was stored (`appdb/access_requests.py`) (PR #208).** A key whose list still held the entry with a trailing invisible character (written before this release) is matched on the cleaned entry too, so the approval no longer finds nothing to remove.

### Added

- **A preflight check, `Column restrictions name real columns` (`scripts/verify_deployment.py`, `core/config_doctor.py`) (PR #208).** It reads the keys of the key file or `API_KEYS_JSON` and of the application database, as the server does, and FAILs, naming the key id and each entry (`ascii()` form), when a plain `denied_columns` entry matches no column of `schema.yaml`. It SKIPs without any key, with an invalid key configuration or without a readable `schema.yaml`. The server is not stopped by the entries (see Fixed). It reads files and the key set only, so the admin panel's preflight card runs it. The preflight has twenty-five checks with one data source (twenty-four in 6.15.0).
- **Look-alike columns, as advice in the same check (`security/column_policy.py`) (PR #208).** The same personal number lives under many names in a warehouse (`NationalID`, `NationalCode`, `CustomerNationalID`). When every entry matches, the PASS detail starts `advice:` and lists, for each denied plain name, the columns that probably hold the same data and are not denied: a column whose name contains the denied name, or contains its stem with another identifier ending (`ID`, `Code`, `No`, `Num`, `Number`). It counts and names columns, never values, ignores the denied name itself and what the key already denies, compares the denied name without any format character inside it, ignores names shorter than five characters (`ID` sits inside thousands), and never fails: it is a guess from spelling, so read each name and add the ones that hold the data.
- **`security.column_policy` gains `normalize_denied_entry`, `normalize_denied_entries`, `unmatched_bare_entries`, `look_alike_columns` and `schema_column_names`; `core.config_doctor.SchemaTable` gains `columns` (the column names, as written) (PR #208).** Documentation: runbook §2 (denied_columns), §3 and §3.1; sections 2.2 and 0.4 of both setup guides; `docs/api-contract-v2.md`.

### Upgrading

- **No setting is renamed or added, no file format changes and every `api_keys` file that starts today still starts (PR #208).** A key whose entries were clean behaves exactly as before, and its cache scope is the same, as is that of a key with a Persian column name that holds a zero-width non-joiner. A key with an entry that normalisation changes now restricts that column, so an answer that used to show it is refused (`denied_column`), and the key's cache partition changes once, so earlier cached results for it are rebuilt on the next request (they were produced under the fail-open list).
- **Run the preflight and read the log once after upgrading (PR #208).** `python scripts/verify_deployment.py` prints `Column restrictions name real columns`: a `[FAIL]` lists each key id and the entries that name no column (a typo, a renamed column, a name with an inner space); correct them in `API_KEYS_FILE` or `API_KEYS_JSON` and restart, or with `PATCH /admin/keys/{key}/acl` for a key of the admin panel. The start-up log has the same entries as WARNING lines, and one more WARNING per key whose entries were cleaned (`denied_columns changed by normalisation`, with each old and new spelling). An entry with nothing visible in it (only spaces or invisible characters) is kept as written, as before, and reported; remove it. Then read the `advice:` line of the same check and decide, for each look-alike column, whether it holds the same data.
- **Panel keys keep what is stored until they are read or written (PR #208).** A stored list with an invisible character is cleaned when the key is read (one WARNING) and rewritten clean the next time the list is saved. `PATCH /admin/keys/{key}/acl` now stores, and returns in `denied_columns`, the cleaned entries (the same list when none had a space or an invisible character).
- **The audit log now says which key asked (PR #208).** `principal_id` is filled on every record of the web UI from this release on. It is still empty on two kinds of record: those written before the upgrade, and the note the model router writes when a task is sent to a remote model provider (`request_id` starting `remote-llm-`; it belongs to no single caller). No other code writes to this audit log.

## [6.15.0] — 2026-10-20

**Issues fixed:** #197 (the first-install order of the setup guide), #198 (aliases lost when the model's answer looped and was cut off).

A setup release for a first install. On a deployment with two SQL Server data sources and no earlier curated configuration, the setup guide could not be followed to a passing preflight: no tool drafted the table and column descriptions the `curated` check asks for, the guide described `datasources.yaml` after the tools that read it and the first API key after the preflight that needs it, and the wizard's alias step left many tables with no alias when the model repeated itself until its answer was cut off. This release adds `scripts/draft_descriptions.py`, puts the install steps in an order that works, keeps the aliases of a cut-off answer, and stops asking for value rules on dimension tables. No setting or file format changes; one pin moves, `certifi`, for a published advisory. The **Upgrading** notes say what to run.

### Added

- **`scripts/draft_descriptions.py` drafts the descriptions `schema.yaml` lacks, with the wizard's model, for review (`schema_data/describe.py`, `schema_data/describe_prompt.py`) (PR #200).** On a first install the inspector and `sync_schema.py --add-tables` describe a table by its own name and a column by `<type> column`, the preflight rightly refuses that, and there was no tool that wrote anything better: a deployment of 76 tables and 1388 columns had to write 1464 descriptions by hand to pass `project_config/ is curated, not a raw draft`. The command finds every table description that is empty, the inspector's placeholder or only the table's name, and every column description that is empty or only `<type> column` (the preflight's own tests), asks the model once per table (the table key, schema and data source, column names and types, the notes already written, the relationships that name the table and its aliases in `entities.yaml`; no row value), and writes `schema.described.yaml` next to `schema.yaml`. Every description a person wrote, every key, flag, comment and the order stay as they were (the script parses its output and refuses to write a file that differs in anything but the drafted descriptions); each drafted line ends with `# REVIEW`. `--language fa|en` (default `fa`), `--tables 'Schema.*'`, `--resume`, `--force`, `--dry-run`, `--max-tokens`, and `--samples` (see below). A table the model fails on is recorded and the run goes on; the summary gives the counts and lists the failed tables, and the output is written after every table. `schema.yaml` itself is never overwritten, and an existing output is not replaced without `--resume` or `--force`.
- **The file starts with `# AUTO-GENERATED by draft_descriptions.py`, a draft header the preflight recognises (`core/config_doctor.py`, `scripts/verify_deployment.py`) (PR #200).** `project_config/ is curated, not a raw draft` fails on a file that still carries it and passes once the descriptions are real and the line is deleted, the documented way to say a draft was reviewed. `core.config_doctor` gains `DESCRIBE_DRAFT_HEADER` and `DRAFT_HEADERS`, the per-description tests `is_placeholder_description`, `is_name_only_description` and `is_type_only_note` (the counters now use them), `read_schema_text`, and the aliases of `read_entities`.
- **`--samples` for the descriptions, with the wizard's rules (`core/value_shapes.py`) (PR #200).** Off by default, so no row is read or sent. With it, up to five distinct values of each label column (a name, title, kind, status, type, category, class, label, caption or group, of at most 50 distinct values) are shown to the model beside the column; a column named like personal or contact data is never read, and no date, number or code, phone number, address, masked value or error text is shown. `core.value_shapes.label_values` is the part of the wizard's `--value-synonyms` step that decides this, now shared by both.
- Documentation: runbook §2.4 (the command), the first-install list, §2.1, §2.3, §3.1 and §16.3; sections 0.3.4 (the command) and 0.3.5 (the first key) and the install list, 0.4 and the checklist of both setup guides; the README scripts list (PR #200).

### Changed

- **The `project_config/ is curated, not a raw draft` failure now gives the next step for a first install as well as for an upgrade (`scripts/verify_deployment.py`) (PR #200).** It said to restore the curated files from a backup or version control, which does not apply when there are none. It still says that for an upgrade, and for a first install tells the operator to draft the descriptions with `python scripts/draft_descriptions.py`, read `schema.described.yaml`, replace `schema.yaml` and delete its header line. The same two cases are in runbook §3.1 and in section 0.4 of both guides.
- **`setup_project.setup_wizard_llm` takes `mock_fallback` and `sends` (`setup_project.py`) (PR #200).** `mock_fallback=False` raises `WizardLLMUnavailableError` instead of continuing with the mock model, and `sends` names what the caller sends in the notice about a remote endpoint. The wizard's own behaviour and output are unchanged. `draft_descriptions.py` uses it, so it takes the endpoint, the key and the `LLM_ALLOW_REMOTE` rule exactly as the wizard does.
- **`generate_with_meta` accepts a per-call `max_tokens` (`llm/base.py`, `llm/providers.py`) (PR #199).** `OpenAIBackend.generate` and `generate_with_meta` take an optional keyword-only `max_tokens` that replaces `LLM_NUM_PREDICT` for that one request; nothing is remembered between calls, and the SQL path passes none, so its request is unchanged. A backend that cannot apply it (`LLMBackend`'s default, `MockBackend`) ignores it.

### Fixed

- **The install order in both setup guides and the runbook did not work for a first install (PR #200).** It described the databases (`datasources.yaml`) after the tools that read them, put the first API key after the section that needs it, and sent the operator to the preflight with no way to pass its curated check. The order is now: prerequisites, dependencies, `.env`, `datasources.yaml` (every source, passwords in `.env`; with one database the file is optional, since `DB_CONNECTION_URL` and `DB_PASSWORD` are the one source `default`), the template, the `schema.yaml` draft, `sync_schema.py --check`, the descriptions, the wizard, the per-source settings, the first admin key and `api_keys.json`, and only then the preflight. Part 0 of the guides gets a section 0.3.5, "The first key", before the preflight section 0.4, so a reader going top-down meets key issuance first. The template is copied into an existing `project_config/` (`cp -r project_config.example/. project_config/`), because `datasources.yaml` is written before it.
- **A table no longer ends with no aliases because the model looped until it was cut off (`llm/providers.py`, `llm/wizard_llm.py`, `setup_project.py`) (PR #199).** Behind an OpenAI-compatible gateway, a Qwen3-class model answered the alias prompt for many tables with `{"aliases": ["a", "b", "a", "b", ...` repeated until `max_tokens` stopped it. The JSON never closed, and `WizardLLM.generate` "repaired" it by sending the truncated text back, which looped again: `Could not extract valid JSON after retry`, 0 aliases, although the first strings were usable. The alias call (`WizardLLM.generate_aliases`) now keeps every complete string of the cut-off list (`salvage_string_list`, which salvages only a flat list of strings and never a half statement, a rule or an example), deduplicates and caps it with `tidy_aliases` as before, and only when nothing can be salvaged asks once more with the original prompt plus a stricter instruction; the truncated text is never sent back. The alias prompt now says at most 8 distinct aliases, no repeats and no reordered repeats, stop after the list. The call carries its own `max_tokens` (320; see Upgrading) so a loop is cut after a few hundred tokens, not after `LLM_NUM_PREDICT`. No `frequency_penalty` or `presence_penalty` is sent: servers differ in whether they accept them, they would have to be reserved in `LLM_EXTRA_BODY` and in the `llm` status block, and a repeat penalty also penalises the legitimate repeated tokens of a multi-word Persian alias.
- **The wizard says when a list was salvaged or when nothing came back (`setup_project.py`) (PR #199).** The step line reads `3 aliases (salvaged 3 from a truncated answer)` or `0 aliases (no usable answer: cut off at the token limit, ...)`, step 3 and the final summary count and name the tables concerned, `.setup_log.json` records them (`aliases_salvaged`, `aliases_failed`), and the error keeps a short excerpt (80 characters) of the last answer and says what was tried; it is no longer the only signal.
- **Dimension and lookup tables are no longer called fact tables without a measure (`setup_project.py`) (PR #199).** The summary said `N fact table(s) got no rule because they have no numeric column that is not a key` and listed tables of a schema named `..._Dim` and tables of a key and a title. The inspector classes a table by its name first, so `OrderStatus` is a fact table (`order`). Step 4 now leaves out, and asks the model nothing about, a table that is classed `dim`, whose schema is named like a dimension (`dim`, `dims`, `dimension`, `lookup`, as a word of the name), or that is a lookup in shape: one key column, one or two non-numeric descriptive columns and no relationship starting at it. The summary names those apart (`N table(s) classed as fact by name were left out because they look like dimensions`) and keeps the old line for the fact tables that really have no measure. A dimension that had a numeric non-key column, such as `SortOrder`, used to be asked for a value rule.

### Security

- **`certifi` 2026.10.19 (`requirements.lock`).** `pip-audit` reports GHSA-6q48-6c7w-9gw2 against the pinned `certifi` 2025.11.12, the CA bundle `requests` uses for outbound HTTPS (the model endpoint, when it is reached over HTTPS). The fixed release is pinned; no code changes.

### Contributors

- **The CI test legs upgrade pip with `python -m pip` (`.github/workflows/ci.yml`).** pip 26.2.2 refuses, on Windows, to replace its own running executable when called as `pip`, which failed every Windows leg before a test ran.

### Upgrading

- **No setting is renamed or added and no file format changes (PRs #199 and #200).** `requirements.lock` moves one pin, `certifi` (see Security): run `pip install -r requirements.lock`. `scripts/draft_descriptions.py`, `schema_data/describe.py` and `schema_data/describe_prompt.py` are new, and one preflight message changed. A deployment that is already curated never meets the new command or the new header.
- **A first install can now reach a passing preflight without writing every description by hand (PR #200).** After the `schema.yaml` draft and `python scripts/sync_schema.py --check`, run `python scripts/draft_descriptions.py` (the wizard's endpoint and model; add `--language en` for English), read `project_config/schema.described.yaml`, replace `schema.yaml` with it and delete its first line. The command sends table and column names, types and relationships to the wizard's endpoint, so a remote one is refused unless `LLM_ALLOW_REMOTE=true`.
- **Run the wizard again for the tables that got no alias, and drop the rules it wrote for dimension tables (PR #199).** Run it over the same output folder: the merge keeps what you reviewed, and the step line of each table that had none says how many aliases were salvaged. A `business_rules.yaml` from an earlier run may hold value or volume rules for dimension or lookup tables, which the wizard no longer drafts; delete them. Copy the files you accept into `project_config/` yourself, as before.
- **The alias call has its own `max_tokens` (PR #199).** It is 320, or `LLM_NUM_PREDICT` if that is lower. It is not sent at all when `LLM_NUM_PREDICT` is above its default of 512 or `LLM_EXTRA_BODY` is set: both are how the setup guide (section 2.9) makes room for a reasoning model, whose thinking counts against `max_tokens`, and a small cap would leave it no room to answer. Such a deployment still gets the salvage, the stricter prompt and the single stricter retry. The 320 is an estimate from character counts (a worst-case Persian answer of eight phrases and a sentence is 386 characters), not a token count of your model: if your tokenizer is unusually expensive for Persian, raise `LLM_NUM_PREDICT` above 512 to lift the cap.

## [6.14.2] — 2026-10-20

A patch release for the drafts the setup wizard wrote. Run over two SQL Server data sources and the inspector, the wizard of 6.13 and 6.14 produced drafts that passed `verify_deployment` and answered almost every question about one source with `OUT_OF_SCOPE`: two entities for nearly every table, `aliases.yaml` full of raw values from the rows, an empty `examples.yaml`, value rules on dimension tables and a 22-alias list. This release fixes each of them, builds no synonym from a row value unless asked, and teaches the preflight (now twenty-four checks, two of them new) to recognise the drafts that slipped through; an installation that ran that wizard has files to check, and the **Upgrading** notes say which.

### Fixed

- **One entity per table, and the keys of `schema.yaml` (`setup_project.py`, `schema_data/registry.py`, `core/wizard_drafts.py`) (PR #194).** An `entities.yaml` the inspector had drafted (bare keys, `schema:` where the wizard read only `schema_name:`) was read as a second schema of every name, so every name looked shared, every table was written `Schema.Table` and the bare entry survived beside it, without aliases; `schema.yaml` keys a table whose name is unique by its bare name, so the qualified entity matched no key and retrieval lost the table. The wizard now reads the schema from `schema_name:`, from `schema:`, from the key's qualifier or from the one table of that name it found; an earlier entry whose key is the table's bare name or `Schema.Table` is the same entity and is merged; two earlier entries that end up on one key become one (aliases united) in `entities.yaml`, `aliases.yaml` and `business_rules.yaml`, and the summary lists them. Keys are spelled by one function, `schema_data.registry.table_key_for`, shared by the sync engine, the inspector and the wizard, so a name with a space, a dot or a non-Latin letter is the same string everywhere (`[Price List]`, `ref.[Run.Log]`).
- **No synonym is built from a row value by default (`setup_project.py`) (PR #194).** `aliases.yaml` `synonyms` held hundreds of sample values mapped to table names: dates, times, numeric and national-id-like numbers, phone numbers, addresses, IP addresses, partly masked names, error messages and hashes, in the prompt and in a versioned file. The wizard writes none; `--value-synonyms` brings back the safe part (see Added).
- **`examples.yaml` is no longer written empty without a word (`setup_project.py`) (PR #194).** Step 5 built its schema summary from the first 20 tables of the combined snapshot, whichever source they were in; swallowed any model error into a log line; counted items it later dropped for lacking a question or SQL; and a `--resume` reused an empty stored result and left the empty file in place. It now asks once per data source about that source's tables (fact tables first), accepts the answer however the model nested or named it, drops anything that is not a question with a `SELECT`/`WITH` statement and says how many, asks once more with a smaller request when nothing is usable, never reuses an empty result, rewrites an `examples.yaml` that holds no example, and when the model gave nothing says so in the step, in the summary (`WARNING: examples.yaml has no example from this run`) and in the file (`# NO EXAMPLES`).
- **No value or volume rule for a table that has no measure (`setup_project.py`) (PR #194).** A rule such as `Value column: ID. Volume column: ID.` was drafted for dimension tables. A fact table now gets a rule only when it has a numeric column that is not a key (not the primary key, a foreign key or join column, or a name ending `ID`, `Key`, `Code`); the prompt lists those columns, a column the model names that is not one of them is dropped, and the tables without one are listed in the summary.
- **Alias lists are cleaned (`core/wizard_drafts.py`) (PR #194).** One table got 22 aliases that were concatenations of the same three words. Aliases are compared after folding case, Persian and Arabic letter forms, digits and punctuation; repeats are dropped; a list of more than 8 loses the aliases whose words all belong to a shorter alias of the same table and then keeps the shortest 8. What an operator types is kept as typed.
- **The preflight caught the 6.13 drafts only by their header line (`core/config_doctor.py`) (PR #194).** `project_config/ is curated, not a raw draft` passed a sync-format `schema.yaml` whose header had been deleted, because its descriptions are the table's own name (`description: 'Sales_Fact.Trade'`) and its column descriptions `bigint column`, not the 6.11 placeholder. It now fails when most table descriptions are only the key, the bare name or `Schema.Table`, or most column descriptions are only `<type> column`, with the counts; and the wizard's own header, `# AUTO-GENERATED by setup_project.py`, counts as a draft header like the inspector's.

### Added

- **`setup_project.py --value-synonyms` (`core/value_shapes.py`) (PR #194).** Writes `aliases.yaml` synonyms from the values of label columns (a name of `name`, `title`, `kind`, `status`, `type`, `category`, ...) that hold at most 50 distinct values, at most 20 per table. Never taken: dates and times, numbers and codes, hashes, phone numbers, e-mail and IP addresses, URLs, addresses, masked values, error text, anything over 40 characters, any column named like personal or contact data (`FirstName`, `Family`, `Mobile`, `Phone`, `Address`, `Email`, `NationalCode`, ...) and the `Name` or `Title` of a table of people (`Customer`, `Employee`, `User`, ...). It reads up to 51 distinct values of each such column. The trade-off is documented: fewer words reach retrieval, and no row value reaches the prompt or git.
- **Two preflight checks, `Aliases hold no raw data values` and `Entities match schema keys` (`scripts/verify_deployment.py`, `core/config_doctor.py`) (PR #194).** The first fails when more than 5 `synonyms` keys have the shape of a date, a phone number, an IP address, a masked value, a long digit run or an address, and reports the count per shape (`<date>: 20, <phone>: 9`), never a key. The second fails when an entity key is not a table key of `schema.yaml`, when two entities stand for one table, or when a `ring_aliases` key names a table by something other than its key; it prints the key to use. Both read files only, so the admin panel's preflight card runs them. The preflight now has twenty-four checks with one data source (twenty-two before).
- Documentation (PR #194): runbook §2.2 (the option, the steps, the one-entity-per-table and no-row-values rules and the trade-off), §2.3 and §3 and §3.1 (the two checks and the widened curated check); the same in both setup guides (sections 0.3.1 and 0.4); `docs/design/TABLE-NAMES.md`; and the upgrade checklists, a step in runbook §17 and in part 4 of both setup guides.

### Changed

- **The alias prompt no longer shows the model raw values (`setup_project.py`, `core/value_shapes.py`) (PR #194).** The sample values put in the per-table alias prompt skip every column named like personal or contact data and every value shaped like a date, a number or code, a phone number, an address, a masked value or error text; short statuses and kinds stay, because they are what makes an alias useful.
- **A re-run drops the value-shaped synonyms of an earlier `aliases.yaml` (`setup_project.py`) (PR #194).** New option `--keep-existing-synonyms` keeps them.
- **The shipped `project_config.example/entities.yaml` names real keys of the template's `schema.yaml` (PR #194)** (`Customer`, `Currency`, `OrderStatus`), so a fresh install that copies the template does not fail `Entities match schema keys`.
- **Table keys are spelled by `schema_data.registry.table_key_for` everywhere (PR #194).** The single-database inspector's drafts used to bracket only the parts that needed it (`[ref].[a.b]`, `Price List`); they now write what the sync engine writes (`ref.[a.b]`, `[Price List]`), as `docs/design/TABLE-NAMES.md` describes.
- **Step 5 makes one model call per data source (PR #194)**, not one for the whole snapshot; a single database is still one call.

### Upgrading

- **If you ran the wizard of 6.13 or 6.14, check what it wrote before you rely on it (PR #194).** Run `python -m scripts.verify_deployment` on the folder you deployed (or the drafts, with `PROJECT_CONFIG_DIR` pointing at them). `Entities match schema keys` lists the entities that name no table of `schema.yaml` and the tables that have two; `Aliases hold no raw data values` counts the synonyms copied from rows; `project_config/ is curated, not a raw draft` fails on a file that still has the wizard's header line, or on a `schema.yaml` of `sync_schema.py` descriptions and `<type> column` notes. Fix `entities.yaml` and `aliases.yaml`: one entity per table under the key `schema.yaml` uses (the aliases of both united), and the same keys in `ring_aliases`. Running the new wizard again over the same output folder does the merge for you, and prints what it merged.
- **Remove the value synonyms: run the new wizard over the same output folder, or delete them by hand (PR #194).** The merge now drops every `synonyms` key of an earlier `aliases.yaml` that is a date, a time, a phone number, an IP address, a masked value, a long digit string, an address, a hash or an error message, and counts them per shape in the summary and in `.setup_log.json`, never listing a key (`--keep-existing-synonyms` keeps them). A key that is a person's name is not recognisable by its shape: read the file too, and delete such keys. The values were in the prompt, and in git if the file was committed; treat a committed copy as having published them. Copy the cleaned `aliases.yaml` into `project_config/` yourself; the wizard never writes there on its own.
- **A preflight that passed on 6.14.1 can fail on this version (PR #194).** It is the same check telling you a file is a draft, now able to see it without the header line: restore the curated files from a backup or version control and add new tables with `python scripts/sync_schema.py --add-tables 'Schema.*'`. A file made mostly of tables added with `--add-tables`, none of them described, now fails too; describe them. The preflight has twenty-four checks with one data source (twenty-two in 6.14.1); the two new ones read files only, so the admin panel's preflight card runs them too.
- **No setting is renamed or added, and `requirements.lock` has no new pin (PR #194).** `core/value_shapes.py` is a new module, and `setup_project.py` has two new options, `--value-synonyms` and `--keep-existing-synonyms`. One template changed: `project_config.example/entities.yaml` now names real keys of the template's own `schema.yaml` (`Customer`, `Currency`, `OrderStatus`) instead of `Entity1` to `Entity3`. A copy of the old template that was never filled in now fails `Entities match schema keys`, because `Entity1` is not a table of `schema.yaml`; a file you wrote yourself is not touched. The summary of a wizard run gains lines for the rules left out, the duplicates merged, the examples and the synonyms.

## [6.14.1] — 2026-10-19

A patch release for one confusing failure: a browser page that cannot use the API while every check passes. In 6.14.0 the preflight, the CLI and `curl` all passed and the web UI still showed three red status lights, because the page had been opened at an address (`http://localhost:8077`) that `CORS_ALLOWED_ORIGINS` did not list (it held `http://172.16.101.42:8077`), and the server logged only `OPTIONS ... 400`. The server now names the refused origin in its log, once per origin, with the `.env` line that allows it; a new preflight check, `Web UI origin is allowed by CORS`, tests an address before anyone opens the page; and the web UI and the admin panel say at the top of the page, once and in Persian, when the API address they were given is not the API (typically the static file server's own address) or cannot be reached or was refused. The CORS policy itself is unchanged: no wildcard, no new default, and a refused preflight is still answered `400`.

### Fixed

- **A refused CORS origin is named in the server log (`api/cors_diagnostics.py`) (PR #192).** In 6.14.0 `verify_deployment` passed fully yet the web UI showed all three status lights red, and the log held only `OPTIONS /health`, `OPTIONS /v2/sessions` and `OPTIONS /v2/access-requests`, each `400 Bad Request`: the allowlist held `http://172.16.101.42:8077` and the page had been opened as `http://localhost:8077`. Nothing said so. A small ASGI layer registered outside `CORSMiddleware` now logs one `WARNING` per distinct rejected origin for the life of the process, for a preflight and for an ordinary request alike: `CORS: origin '<origin>' is not allowed (allowed: <list>). Browsers show this as 'Failed to fetch'. To allow it add this line to .env and restart: CORS_ALLOWED_ORIGINS=<current>,<origin> -- or open the page with exactly an allowed address (scheme, host and port). Logged once per origin.` At most 64 origins are remembered, then one line says further ones are not logged; the origin is stripped of control characters, cut at about 200 characters and never put into a response. The CORS policy is unchanged (no wildcard, no new default; the preflight is still answered `400`), and a request to the server's own address (`Origin` equal to `Host`) is same-origin and not reported.
- **The web UI says what is wrong with the API address, once, in Persian (`web/js/diagnosis.js`) (PR #192).** The status page showed red lights and a tooltip nobody reads; every admin card showed its own 404 when the address saved in the browser was the static file server's. A failed call to the configured address is now classified as a wrong address (it is the page's own origin, or `GET <address>/health` answers a 404, an HTML page or JSON that is not the API's health response) or as refused-or-unreachable (the call got no response; a browser cannot tell the two apart, so the message names both with their remedies and the page's own origin as the value to add to `CORS_ALLOWED_ORIGINS`). The analyst status view and the admin panel show it as one banner at the top; each admin card then says only «به API نرسید؛ پیام بالای صفحه را ببینید.» instead of repeating it. A 401, 403, 429 or 5xx is the API answering and keeps its own message, and the capability-specific 403 messages are unchanged. The analyst status view also no longer treats a 200 whose body is not the API's health response as a healthy API.

### Added

- **A preflight check, `Web UI origin is allowed by CORS` (`scripts/verify_deployment.py`, `core/cors_origins.py`) (PR #192).** With `VERIFY_UI_ORIGIN` set to the address the UI is opened with (for example `http://172.16.101.42:8077`) it passes when that origin is in `CORS_ALLOWED_ORIGINS` and fails otherwise, printing the `.env` line to add. Without it nothing can be proved, so it only advises (a `[PASS]` line starting `advice:`): when every allowed origin is loopback while `API_HOST` is not, or when exactly one non-loopback origin is allowed, it says to list every address the page is opened with (IP address, host name, `localhost`) and to set `VERIFY_UI_ORIGIN` to check one. It reads settings only (no network, no database), so the admin panel runs it too. The preflight now has twenty-two checks with one data source.
- Documentation (PR #192): runbook §2 (the origin list must hold every address), §3 and §3.1 (the check and `VERIFY_UI_ORIGIN`), §4 and §5 (the new log line), and a new §5.1 table of the two symptoms with the log line and the fix; the same additions in section 2.3.1, section 2.8 and the preflight section of both setup guides (the check count is now twenty-two); `web/README.md`; the CORS block of `.env.example`; and the upgrade checklists, a step in runbook §17 and in part 4 of both setup guides.

### Upgrading

- **No file format and no dependency changes (PR #192).** No setting is renamed or added to `config.py` or `.env`, no file's template changed, and `requirements.lock` has no new pin (`git diff v6.14.0 -- requirements.lock` is empty): run `pip install -r requirements.lock` as for every upgrade, but nothing new is installed. The audit log, the session store, the API's responses and the CORS policy are as in 6.14.0; the server change is one log line, from a layer around `CORSMiddleware` that never answers a request and changes no request or response.
- **A new `WARNING` appears in the server log when a browser origin is refused (PR #192).** One line per distinct origin per process, for a preflight and an ordinary request alike, naming the origin, the allowed list and the `.env` line to add (`CORS: origin '...' is not allowed (allowed: ...). Browsers show this as 'Failed to fetch'. ...`). It means a browser was opened at an address the list does not hold, not that the server is unwell. At most 64 distinct origins are remembered per process; after that one line says further ones are not logged.
- **A new preflight check, `Web UI origin is allowed by CORS`, and the optional `VERIFY_UI_ORIGIN` (PR #192).** The preflight now has twenty-two checks with one data source (twenty-one before). It reads settings only, so it needs no network and no database, and the admin panel's preflight card runs it too. Without `VERIFY_UI_ORIGIN` it never fails: it passes with an `advice:` line when every allowed origin is loopback while `API_HOST` is not, or when exactly one non-loopback origin is allowed. Set `VERIFY_UI_ORIGIN` to the address the page is opened with, as the browser's address bar shows it (`http://172.16.101.42:8077`), to make the check pass or fail on that address; a value that is not `scheme://host[:port]` fails. It is read from the environment of the process that runs the preflight, so on the admin panel it has to be set in the server's own environment to be more than advice.
- **List in `CORS_ALLOWED_ORIGINS` every address the page is opened with; the API's own address is not needed there (PR #192).** The IP address, the host name and `localhost` are three different origins, and so are two ports on one machine. The variable replaces the default list, so write all of them, comma-separated, and restart (for example `CORS_ALLOWED_ORIGINS=http://172.16.101.42:8077,http://my-host:8077,http://localhost:8077`). A request whose `Origin` is the server's own address (a page served by the API itself) is same-origin, needs no entry and is not reported. Run the new check once per address people use.
- **An admin panel whose saved API address points at the static server now says so (PR #192).** When the address saved in the browser (it takes precedence over `web/js/config.js`, and the top bar saves it) is the static file server's, every card used to show its own 404; now a banner at the top says that the address is not the API and each card points to it. Fix it in the address bar at the top of the admin panel (enter `http://<host>:8000` and press «اتصال»), or load the page with `?base=http://<host>:8000`, which takes precedence over the saved value for that load. The analyst page shows the same banner and names `web/js/config.js` (`DEFAULT_BASE_URL`) instead, since it has no address bar.

## [6.14.0] — 2026-10-18

A benchmark track for accuracy, beside the private golden set, and confidence intervals on every figure. `python -m eval.cli bench-import bird` loads the public BIRD mini-dev benchmark (500 questions over 11 SQLite databases, CC BY-SA 4.0) into a database of the operator's own, one schema per database, and writes a golden file in the existing format whose reference results come from the original SQLite files; `bench-run bird` runs it through the production pipeline with the operator's model, once per database with that database's configuration, and merges the reports. It gives development decisions a large, fully visible sample whose full report can be shared because the data is public, and it is never the deployment's accuracy: the report says so in its first line. Separately, every report now prints a Wilson 95% interval and the number of cases for its overall figure and each stratum (a stratum under 20 cases is "not interpretable"), `compare` adds the paired statistics (fixed, regressed, an exact McNemar p-value, a 95% interval of the difference), and a determinism probe can report outcome consistency and mean accuracy over its repeats.

### Added

- **A benchmark track: `python -m eval.cli bench-import bird` (`eval/benchmarks/bird.py`, `bird_import.py`, `bird_config.py`, `sqlite_load.py`) (PR #190).** From the unpacked BIRD mini-dev package it copies each selected SQLite database into one schema of a SQL Server database named after the `db_id` (SQLAlchemy reflection; types mapped from what the data holds, `BIGINT` / `FLOAT(53)` / `NVARCHAR(n)` / `VARBINARY(MAX)`, dates stay text; names kept and quoted; chunked inserts with pyodbc `fast_executemany`; primary and foreign keys added after the data, a key SQL Server refuses is a warning), checks that every table has the source's row count, and writes `golden.jsonl` (and with `--with-evidence`, `golden.with_evidence.jsonl`), overall and per database, an `import_report.json` / `.txt` and a `NOTICE.txt` with the licence attribution. The load takes its own administrator login (`--target-url`, or `env:NAME`); the application then reads the database with its read-only login. A non-empty schema is refused, before anything is written, unless `--replace` is given; `--skip-load` re-checks schemas already loaded. `--dbs a,b` selects databases.
- **The reference result comes from the original SQLite file, and the copy is shown to be faithful by its data (PR #190).** Each case's expected rows are the result of the dataset's own SQL on its own `.sqlite` (opened read-only) and are compared by execution accuracy, so no translation can make a question wrong. After the load every table is compared with its source by content, not only by row count: rows are reduced to a canonical form (integers as integers, floats to 12 significant digits, text exactly with case and trailing spaces, blobs by hash) and the sum of the row hashes is an order-independent digest computed in Python on both sides in chunks. If every table matches, no question is excluded for a doubt about the load; a table that differs is reported with its columns, and only the questions whose reference SQL reads it are excluded, with the reason `load_fidelity`. Other exclusions, all kept in the file and listed one by one in the import report: `reference_failed`, `reference_empty`, `reference_too_large`. Cases are tagged `bench:bird`, `db:<db_id>` and `difficulty:<x>`; the evidence hint is off by default and, with `--with-evidence`, appended to the question under a marker for an ablation.
- **The target database must be case- and accent-sensitive (PR #190).** SQLite compares text case-sensitively; in a default SQL Server database `WHERE name = 'abc'` also matches `'ABC'`, so a correct answer could differ from the reference. The import reads `DATABASEPROPERTYEX(DB_NAME(), 'Collation')` and refuses a database that is not (a `CS` and `AS` collation, or `BIN` / `BIN2`) before writing anything, naming `CREATE DATABASE ... COLLATE Latin1_General_100_CS_AS_SC_UTF8`; `--allow-ci-collation` imports anyway and the import report and every report header then flag it.
- **A per-query check is a diagnostic, not a filter (PR #190).** Each reference is also transpiled to T-SQL with sqlglot, run on the copy and compared. A difference on identical data is a dialect-semantics difference (integer division, string comparison, a mixed-type column): the case is tagged `fidelity:semantic_difference`, stays in the count by default because the SQLite answer is still the right answer, and is listed; `--exclude-semantic-difference` removes it (reason `semantic_difference`). A reference that cannot be transpiled or fails on SQL Server is tagged `fidelity:unverified`, information only, and never excluded: excluding the queries that are hard to translate would inflate the accuracy.
- **A report gives the accuracy on all scored cases and without the semantic-difference cases (PR #190).** A benchmark report states scored, excluded-by-reason and flagged counts in its header, and the JSON carries both figures under `fidelity`.
- **A `project_config/` per database, and one over all schemas (PR #190).** `--out/<db_id>/project_config/` is the 6.13 inspector's draft (`--include-schemas <db_id>`) read through the application's login, completed with `system_prompt.md` and the policies from `--base-config` and empty `business_rules.yaml`, `examples.yaml`, `metrics.yaml` and `retrieval_hints.yaml`; `--descriptions` adds BIRD's column descriptions to `schema.yaml` (a curated configuration; without it the raw draft). `--out/_combined/project_config/` covers every schema.
- **`python -m eval.cli bench-run bird` runs and merges (PR #190).** It starts `eval.cli run --live --reference stored` once per database, each process with `PROJECT_CONFIG_DIR` set to that database's configuration, then pools the reports into `runs/<label>/report.json` and `report.txt`. `--mode combined` is a separate measurement over the combined configuration, labelled as such; `--evidence` runs the hint variant; `--label` names the folder; arguments after `--` go to every `run`.
- **A report of a benchmark says so (PR #190).** A report whose cases all carry the same `bench:<name>` tag begins with `BENCHMARK RUN (<name>) ... This is NOT this deployment's accuracy`, and the merged report names the mode, the evidence setting and the configuration; `EvalReport.benchmark` carries it in the JSON.
- **A golden case can be `excluded`, with an `exclusion_reason`, and can be matched by execution (`eval/models.py`) (PR #190).** `status: "excluded"` is skipped by the runner like the pending statuses and `run` prints `Not scored: N of M case(s) -- excluded ...` with the reasons; `match: "execution"` makes `--reference stored` compare the result with the recorded `expected_rows` by execution accuracy (column names ignored, row order only when the reference sorts with `TOP`) instead of by fingerprint. Both fields are optional and default to the previous behaviour.
- **Wilson 95% intervals on every report (`eval/stats.py`) (PR #190).** The text report prints the interval and `n` under the overall accuracy and on every per-tag and per-source line, and a stratum with fewer than 20 cases is marked `n<20: not interpretable`; the `--out` JSON and the share file's `summary` carry the same under `intervals`.
- **`compare` reports the paired statistics (PR #190).** On the cases both rounds scored: fixed, regressed and unchanged counts, the exact two-sided McNemar p-value on the discordant pairs, and a Newcombe 95% interval of the accuracy difference, in the text and under `paired` in `--json`. No new dependency: the binomial tail is summed directly. An older report compares as before.
- **The determinism probe can report outcome consistency and mean accuracy over its repeats (PR #190).** With `--live --determinism` each repeat is also executed and judged like the run; the report adds the share of cases right on every repeat or wrong on every repeat, and the mean accuracy over the repeats.
- Documentation (PR #190): runbook §18.7 (the benchmark track: why it exists and why its number is not the deployment's accuracy, licence and download, loading with an administrator login and reading with the application's, import, per-database and combined runs, the evidence ablation, what to send back), §18.3 and §18.6.2 (intervals, the paired statistics, repeats), and a section 5.6 in both setup guides.

### Changed

- **`eval.verify.frame_to_rows` moved to `eval/rows.py` (PR #190)** so the benchmark importer can record rows without a `project_config/`; `eval.verify` still exports it.

### Upgrading

- **The server does not change, and nothing has to be done by a deployment that does not use the benchmark track (PR #190).** The code that changed is the `eval` package (the accuracy harness, run from the command line) and its tests; `api/`, `core/` and the other server packages are as in 6.13.0 apart from the version number, so the answers the server gives, the audit log and the admin panel are unchanged. No setting, dependency or `requirements.lock` pin changed (`git diff v6.13.0 -- requirements.lock` is empty): run `pip install -r requirements.lock` as for every upgrade, but nothing new is installed.
- **Reports gain intervals and paired blocks, all of it additive, so a report from 6.13.0 or earlier still loads and compares (PR #190).** The text report prints a Wilson 95% interval and the number of cases under the overall accuracy and at the end of every per-tag and per-source line; the `--out` JSON gains the keys `intervals`, `benchmark` and `fidelity` (the last two are `null` for a golden set of your own); the share file's `summary` gains `intervals` and its `format_version` stays 1; `compare` prints a paired block and `compare --json` returns it as `paired`. Nothing was removed or renamed. A baseline, `--out` report or share file recorded on 6.13.0 or earlier is read as before: `--baseline` gates against it and `compare` takes it as either side, the paired block being computed on the cases both rounds scored; there is no need to record the baselines again. A script that reads the per-tag or per-source lines of the text report by column has to allow for the interval now at the end of each line; the JSON is the stable source.
- **The golden set has a new status, `excluded`, and a new field, `match`; both are optional and your own set never uses them (PR #190).** An `excluded` case (with a required `exclusion_reason`) is skipped by the runner like the pending statuses, and `match: "execution"` makes `--reference stored` compare the recorded `expected_rows` by execution accuracy instead of by fingerprint. Only `bench-import` writes either; `harvest_golden.py`, `golden_sheet.py` and `verify` do not, and a case without them is read and scored exactly as before, so an existing `golden.jsonl` needs no edit. Two things are visible: `run` now prints a `Not scored: N of M case(s) -- ...` line under the report whenever the file holds a case that is not `active` (a pending one included, which was skipped without a word before), and a golden file the tools rewrite now carries `"exclusion_reason": null` and `"match": "fingerprint"` on every line, which 6.13.0 also reads, because it ignores keys it does not know.
- **`--live --determinism` now also executes and judges each repeat (PR #190).** The probe used to call the model only. It now runs every repeat's SQL through the guard and the read-only executor, as `run` does, so one probe sends up to cases times repeats (3 by default) extra read-only queries to the data sources, and its report adds the outcome consistency and the mean accuracy over the repeats. Size a probe on a large golden set for the warehouse's quiet hours.
- **The benchmark track is optional, and nothing runs it unless you call `python -m eval.cli bench-import` or `bench-run` (PR #190).** If you decide to use it (runbook §18.7):
  - BIRD mini-dev is not part of this repository. Download it yourself and keep it, the loaded database and the files `bench-import` writes outside the checkout.
  - The load needs a separate administrator login, `--target-url` (or `env:NAME`, which keeps the password off the command line), that may create schemas and tables in a database of its own; the application reads that database with its normal read-only login, and the application's login must never be given the right to create anything. `--datasource` without `--target-url` would load through the application's login, and the command prints a warning saying so.
  - The database must be case- and accent-sensitive: a collation with `CS` and `AS` in its name, or a `BIN` / `BIN2` collation (create it with `COLLATE Latin1_General_100_CS_AS_SC_UTF8`). `bench-import` reads the collation and refuses any other before it writes anything; `--allow-ci-collation` imports anyway, and the import report and every report header then carry a warning.
  - Its figure is not the deployment's accuracy. The questions are public, English and about other domains, and the report begins `BENCHMARK RUN (bird) ... This is NOT this deployment's accuracy`. It is a second instrument for development decisions and does not replace the golden set built from your own questions (runbook §18.1).

## [6.13.0] — 2026-10-16

Which schemas the setup tools skip is now one rule, kept in configuration, and the two tools that draft a project configuration read every data source the way the application does. `EXCLUDE_SCHEMAS` in `.env` and `exclude_schemas:` per data source name the schemas that `database.schema_inspector_cli`, `scripts/sync_schema.py` and `setup_project.py` do not scan, on top of the system schemas, and a preflight check reports a `schema.yaml` table that sits in one. The inspector and the wizard now read each source of `datasources.yaml` through the application's own engine, so a draft lists what the application can query; the inspector writes the format `sync_schema.py` writes and puts no sample value into `schema.yaml`, `entities.yaml` or `aliases.yaml`. The wizard writes to `project_config_draft/`, merges a second run into the first instead of replacing it, keys tables of the same name in different schemas separately and can record the schemas an operator left out in `datasources.yaml`. The setup guide now exists in English beside the Persian one, and a test keeps the two in step.

### Added

- **One rule says which schemas the setup tools do not scan, and it is kept in configuration (`database/schema_filter.py`) (PR #184).** A schema is excluded when its name matches, without regard to case, any of: the system schemas (`sys`, `INFORMATION_SCHEMA`, `guest`, `cdc`, the nine fixed `db_*` role schemas by exact name, `pg_catalog`, `pg_toast`, `pg_*`), the new `EXCLUDE_SCHEMAS` setting (comma-separated `fnmatch` patterns, every data source, so a deployment with only `DB_CONNECTION_URL` has the same control), the data source's new `exclude_schemas:` list in `datasources.yaml` (validated as non-empty strings), and the patterns of one command line. A user schema called `db_sales` is still scanned. `schema_is_excluded` moved there and is still importable from `database.schema_inspector`.
- **`python -m database.schema_inspector_cli` reads every data source through the application's own engine and writes drafts in the format `sync_schema.py` writes (PR #184).** With no `--db-url` it inspects each source of `datasources.yaml` (or the implicit `DB_CONNECTION_URL` source) with the login the server uses, so a draft lists only what the application can query. `--datasource NAME` (repeatable) narrows the run. `schema.yaml` is built by `schema_data.sync.draft_schema_text`, the code behind `--add-tables`: `datasource:` lines with several sources, a table with the same `schema.table` in several sources written once as `datasource: [a, b]`, `Schema.Table` keys only for names shared across schemas (counted across all selected sources). `entities.yaml` uses the same keys.
- **The inspector drafts relationships when the database declares no foreign key (PR #184).** Declared keys and the naming rules of `sync_schema.py`'s proposals (`<Table>_ID`, `<Table>ID`) both feed the draft; a guess carries the comment `# INFERRED from column names (not a declared foreign key) -- REVIEW before use`, a declared key none.
- **Staging and change-capture tables (`STAGING_PATTERNS`) are left out of the inspector's draft by default; `--include-staging` keeps them (PR #184).** The end-of-run summary says how many tables and schemas each rule skipped: system schemas, configured exclusion, `--exclude-schemas`, `--include-schemas`, `--exclude-tables`, staging.
- **`--sample-values` (with `--sample-rows N`, default 10) writes real column values to `sample_values.review.yaml` in the draft directory (PR #184).** The file is readable by its owner only (`core/fileperms.py`); nothing loads the file and `--dry-run` never prints it.
- **`sync_schema.py` honours the exclusion (PR #184).** The "in the database, not in `schema.yaml`" report and `--add-tables` skip tables in excluded schemas (`--add-tables '*'` adds only tables outside them), and the report says how many were skipped, per schema. `--exclude-schemas PATTERNS` (repeatable) adds patterns for one run. A table `schema.yaml` already lists is synced exactly as before, and an excluded table never makes `--check` fail.
- **A preflight check, `No table in an excluded schema`, reports a `schema.yaml` table whose schema every one of its sources excludes (PR #184).** It follows the other doctor advice: a `[PASS]` line starting `advice:` with the count and the first keys (never a value) and the two fixes, remove the table or narrow the exclusion; `[PASS]` when none; `[SKIP]` when nothing is excluded beyond the system schemas and no table matches. File reads only.
- **The setup wizard describes every data source of `datasources.yaml`, through the application's own login (`setup_project.py`, `database/setup_sources.py`) (PR #187).** With a `datasources.yaml` and no `--db-url`, step 1 opens each source with `database.connection.get_engine` and step 2 lists, plans, inspects and keys the tables of all of them together; `--datasource NAME` (repeatable) narrows the run, and `--db-url` keeps reading one database with the login in the URL. A bare table name shared by several schemas, in one source or across sources, is keyed `Schema.Table` (the inspector's and `sync_schema.py`'s rule), and a `schema.table` present in two sources is one entity. With several sources configured every `entities.yaml` entry carries `datasource: [sales, reports]`, written by the logic of `database/schema_draft.py` (`listed_datasources`); a later run for one source adds to an earlier entry's list instead of replacing it. `SchemaInspector.for_engine(engine)` reads through an engine it does not own and leaves it open. The resume fingerprint includes the names of the sources read and the sources of each table. The inspector's source selection is the same code now (`database.setup_sources.configured_sources`).
- **The wizard applies the configured schema exclusion and says which rule skipped what (PR #187).** Each source's reading plan uses the rule of `database/schema_filter.py`: the system schemas, `EXCLUDE_SCHEMAS`, that source's `exclude_schemas:` and `--exclude-schemas`, so a configured schema is never listed in the question, read or sampled. The skip summary adds the inspector's per-rule lines (`system schemas`, `configured exclusion (EXCLUDE_SCHEMAS / exclude_schemas:)`, `--exclude-schemas`, `left out interactively`, `staging`), one per source with several.
- **The wizard can record the schemas an operator leaves out interactively (PR #187).** It offers (default No; never with `--non-interactive` or `--dry-run`) to add them as the matching source's `exclude_schemas:` in `datasources.yaml` (`database/datasources_edit.py`): a comment-preserving text edit after a `datasources.yaml.bak-<YYYYmmdd-HHMMSS>` copy, re-validated with `validate_datasources_yaml_text` and compared with the original before the file is replaced; on any failure the file is left as it was and the lines to add by hand are printed. Without `datasources.yaml` (or with `--db-url`) it prints the `EXCLUDE_SCHEMAS=...` line for `.env` and never edits `.env`.
- Documentation (PR #184): runbook §2.3 (the inspector, rewritten), §3.1 (the new check), §16.2 and §16.3 (`exclude_schemas:`, the exclusion in the sync); `docs/design/DATASOURCES.md`; `docs/design/TABLE-NAMES.md`; `exclude_schemas:` and `EXCLUDE_SCHEMAS` in the example files.
- **The setup guide is bilingual: `docs/en/getting-started.md`, kept in step with the Persian guide by a test (PR #186).** The step-by-step guide existed only in Persian (`docs/fa/getting-started.md`). The English version has the same sections, numbering, tables and commands: only prose and comments are translated, and a Persian string that the operator sees in the UI (a button label, a card title) stays as it appears, with its English meaning in parentheses. Both files now open with a language switch, and the README links to both. `tests/test_setup_guide_parity.py` compares what survives translation: the heading outline with its section numbers (Persian digits folded to ASCII), the number of fenced code blocks (in total and section by section), the multiset of `--flags` and of `NAME=` environment variables, and the set of documents the guide names; a failure names the first section that differs, and a copy with one heading removed is shown to fail.
- Documentation (PR #188): the setup guides in both languages (`docs/fa/getting-started.md`, `docs/en/getting-started.md`) describe the 6.13.0 setup tools. Sections 0.3.1 to 0.3.3 and the install list say that the wizard and the inspector write drafts to `project_config_draft/`, read every data source of `datasources.yaml` through the application's login and share one schema-exclusion rule with `sync_schema.py`; the preflight section lists the new check (twenty-one checks in all); the fix for `Tables are in their data source` is `sync_schema.py`; part 4 carries the upgrade note about `.setup_log.json`; and an unclear sentence in 5.5 is rewritten. The admin panel is described with its thirteen sections (a key with only `admin` opens 4 and gets 403 on 9), in the guides, `docs/admin-panel-architecture.md` and `web/README.md`.

### Changed

- **The inspector's built-in system-schema filter compared a lower-case `information_schema` and missed SQL Server's `INFORMATION_SCHEMA`, `cdc`, `guest` and the fixed role schemas (PR #184).** All of them are skipped now, in any letter case, by `SchemaInspector.inspect` (used by the setup wizard) as well.
- **The inspector no longer drafts live values into `schema.yaml`, `entities.yaml` or `aliases.yaml` (PR #184).** Nothing in the application reads a `sample_values` key from `schema.yaml`; the `e.g. a, b, c` text drafted into column descriptions went into the prompt, and `aliases.yaml`'s `synonyms:` seeds were sample values. Without `--sample-values` no row is read (only `COUNT(*)` per table, unless `--no-row-counts`), `aliases.yaml` has `synonyms: {}`, and a column description is `<type> column`.
- **The draft `schema.yaml` follows `sync_schema.py`'s spelling (PR #184).** A name that is not a plain identifier is bracketed in its key (`[Price List]`), a qualified key (`sales.Trade`) carries its schema and has no `db_schema` line, and each table has `column_types:`. The inspector's classification and primary-key notes are no longer in `schema.yaml` (`entities.yaml` keeps `classification`, `row_count` and `is_pk`). Views are drafted as well as tables, as `sync_schema.py` does.
- **The setup wizard writes drafts to `project_config_draft/` by default, and refuses the live directory (`setup_project.py`, `core/wizard_drafts.py`) (PR #185).** The default `--output` was `project_config/`: the wizard replaced curated `entities.yaml`, `aliases.yaml`, `business_rules.yaml`, `examples.yaml` and `relationships.yaml` with no backup, which the schema inspector had refused to do since 6.9.3. Writing into the live configuration directory (`PROJECT_CONFIG_DIR`, by default `project_config/`) or `project_config.example/`, compared on resolved paths, now stops the run before the database or the model is touched (exit status 2, one-paragraph message) unless `--overwrite-live` is given; then every file about to be replaced, the setup log included, is first copied to `<name>.bak-<YYYYmmdd-HHMMSS>` next to it and the copies are listed. A file the wizard does not produce is never touched. A `--dry-run` writes nothing and is not refused.
- **A second run merges into the first instead of replacing it (PR #185).** Run once per database, the wizard's later `aliases.yaml` and `entities.yaml` used to replace the earlier ones, so the files shrank. Entries for tables this run did not cover now stay as they are, entries for tables of this run replace the old ones with the same key, examples are de-duplicated by question text and relationships by their two ends; nothing is removed. A table the model gave no answer for does not wipe an earlier entry: empty new aliases, or the placeholder rule written when the model gave none, leave the old entry in place (counted as kept and named in the summary). The summary prints `kept`, `added` and `replaced` per file. `--fresh` starts from scratch (still subject to the protected-directory rule). An existing wizard file that cannot be read stops the run before any model call.
- **The wizard keys entities, aliases, business rules and relationships like the schema inspector (PR #185).** A table is keyed by its bare name when no other table shares it and by `Schema.Table` when several schemas do (`docs/design/TABLE-NAMES.md`), and the model is told the schema-qualified name. A bare entry of an earlier run whose name is now shared is re-keyed in all five files and reported, not duplicated or overwritten by the other schema's table.
- **The wizard can leave schemas and staging tables out before any model call (PR #185).** New `--exclude-schemas PATTERNS` (comma-separated `fnmatch`, case-insensitive, wins over `--include-schemas`) and `--include-staging`. The tables `STAGING_PATTERNS` names are skipped by default, and the summary says how many tables were skipped and why. **A schema that is left out is never read:** step 2 first lists schema and table names (SQLAlchemy's inspector, names only), applies the include, exclude and staging rules and, in interactive mode, shows the schemas with their table counts and asks which to leave out, and only then inspects the remaining schemas with `SchemaInspector.inspect(include_schemas=..., exclude_schemas=..., exclude_tables=...)`, so no column or sample value of a left-out schema is fetched. A schema whose every table is staging is not read; a staging table with a bare name no other table shares is not read; one with a shared name is read and dropped afterwards so a same-named table elsewhere is kept. Non-interactive mode never prompts. The resume fingerprint is taken after the exclusions and includes the table keys.
- **The wizard's skip summary names the rule (PR #187).** `Skipped N of M tables` now also counts system schemas and configured exclusion (`EXCLUDE_SCHEMAS`, `exclude_schemas:`), which `SchemaInspector.inspect` used to drop silently, and `.setup_log.json` records the counts per rule and source (`step2_schema.skipped`).
- **Relationships and the business-rules join hints no longer depend on declared foreign keys (PR #185).** When the database declares fewer than one foreign key per ten tables, `relationships.yaml` also gets the relationships `schema_data/relationship_proposals.py` infers from column names; each carries `basis` and `confidence` fields and a `# REVIEW` comment, a reviewed entry is never replaced by a guess, and the business-rules prompt sees the same join hints.

### Fixed

- **The setup wizard no longer collapses tables of the same name in different schemas (`setup_project.py`) (PR #185).** Entities, aliases and rules were keyed by the bare table name, so `Trade` from four schemas became one entity: a 278-table warehouse produced a draft with 190 entities. Every table now has its own entry.
- **The wizard no longer overwrites curated configuration without a backup, and no longer loses the first database's entries on the second run (PR #185).** See Changed.
- **The warning `scripts/issue_api_key.py` prints for an `admin`-only key counted six sections the key cannot open; there are nine (PR #188).** The admin panel has thirteen sections, `require_admin` serves four of them, and the other nine (maintenance, feedback, schema drift, vocabulary, quality, per-analyst usage, auth failures, keys, access requests) need `operations` and/or `security`. The message names all nine; the key that is issued is unchanged.

### Upgrading

- **`python setup_project.py` now writes its drafts to `project_config_draft/` (git-ignored), not to `project_config/` (PR #185).** Copying the files you accept into `project_config/` is a reviewed step. To write into the live directory (`PROJECT_CONFIG_DIR`, by default `project_config/`) or `project_config.example/`, pass `--overwrite-live`: without it the run stops with exit status 2 before the database or the model is touched, and with it each file about to be replaced, `.setup_log.json` included, is first copied to `<name>.bak-<YYYYmmdd-HHMMSS>` next to it. A script or runbook of yours that ran the wizard and expected `project_config/` to be filled needs `--output project_config --overwrite-live` or a copy step. A `.setup_log.json` written by an earlier version is read as before; `--resume` generates again once, because the schema fingerprint now includes the table keys.
- **Running the wizard again over a directory of an earlier run merges into it (PR #185).** Use `--fresh` for the old start-from-scratch behaviour. Entries keyed by a bare name that is now shared by several schemas are re-keyed as `Schema.Table`; check the summary line `Re-keyed in ...`.
- **The wizard reads `datasources.yaml` when it exists (PR #187).** Run without `--db-url` it describes every source in the file through the application's login and writes `datasource:` lines to `entities.yaml`; give `--db-url` to keep the old single-database run. A deployment with only `DB_CONNECTION_URL` is unchanged.
- **The inspector now uses the application's login by default (PR #184).** Run it without `--db-url` and it reads `datasources.yaml` (or `DB_CONNECTION_URL`) through the engines the server uses, so tables the application cannot query no longer appear in a draft. `--db-url URL` still works for a first install before any configuration exists (one source, read with the login in the URL) and prints one line saying the application's login may see less. The `DATABASE_URL` environment variable is no longer read by `database.schema_inspector_cli`; pass `--db-url` (the wizard's single-database run still falls back to it when there is no `datasources.yaml` and no `DB_CONNECTION_URL`).
- **Drafts no longer contain sample values, and staging tables are skipped (PR #184).** A script that relied on `e.g.` text or `sample_values:` in the drafts must pass `--sample-values` and read `sample_values.review.yaml`; one that wanted staging tables passes `--include-staging`. Both flags are new and nothing else about the existing flags changed (`--exclude-schemas`, `--include-schemas`, `--exclude-tables`, `--no-row-counts`, `--dry-run`, `--output-dir`, `--sample-rows`).
- **One setting and one `datasources.yaml` key are new, both optional and empty by default: `EXCLUDE_SCHEMAS` in `.env` and `exclude_schemas:` per data source (PR #184).** Without them only the system schemas are skipped, which can hide a table `sync_schema.py` used to report in `cdc` or one of the fixed role schemas. A `schema.yaml` that already lists such a table keeps it and is synced as before. Both are read by the setup tools and the preflight only; nothing at run time uses them, so the answers the server gives do not change. An `exclude_schemas:` that is not a list of non-empty strings makes `datasources.yaml` fail to load, like any other invalid key. Both are in the example files.
- **Run `python scripts/verify_deployment.py` after the upgrade to see the new check, `No table in an excluded schema` (PR #184).** It never fails: it is a `[PASS]` line starting `advice:` when a `schema.yaml` table sits in a schema that every one of its data sources excludes (remove the table, or narrow the exclusion), a plain `[PASS]` when none does, and `[SKIP]` while nothing is excluded beyond the system schemas and no table matches. It reads files only. Runbook §3.1 lists it.
- **No `requirements.lock` pin changed and no file format changed (PR #184, PR #185, PR #187).** `git diff v6.12.0 -- requirements.lock` is empty. Run `pip install -r requirements.lock` as for every upgrade, but nothing new is installed.

## [6.12.0] — 2026-10-14

Accuracy rounds get a loop with a reviewer who has neither the model nor the warehouse, and the endpoint they measure is checked for a response cache. The operator runs `eval.cli run --live --reference live --share FILE` on the server and sends that one file; it says not only *that* an answer differed but *how* (too few rows, the right rows in the wrong order, a number two percent off) from a diagnosis computed on the server, and holds no result value. `eval.cli compare` puts two rounds side by side, `eval.cli split` keeps a held-out half nobody tunes to, and `--live` now measures what an analyst actually gets: it generates through the server's own correction loop instead of one model call, and a result cut by the row cap passes as `match_capped` only when a count of the uncapped SQL equals the reference's row count. Separately, a gateway in front of the model can keep a response cache and answer a new prompt with the reply stored for a similar one, which nothing in the reply shows: a setup run wrote one alias list for most of a database's tables, and the same endpoint can serve a valid but wrong SQL for a similar question and make an accuracy round measure the cache. The preflight check `Model endpoint answers each prompt itself` finds it with two tiny requests that differ in one random token, the setup wizard stops when the model gives the identical answer for three tables in a row, and `LLM_CACHE_SUSPECT_MS` warns at run time when a different question comes back with byte-identical SQL faster than the threshold.

### Added

- **A preflight check, `Model endpoint answers each prompt itself`, finds a response cache in front of the model (PR #181).** A gateway between the application and the model server can keep a semantic or prefix response cache and answer a new prompt with the reply stored for a similar one; nothing in the reply says so, and the application cannot tell. The check (`scripts/verify_deployment.py`, `check_model_answers_each_prompt`, right after `OpenAI-compatible model exists`) sends two tiny requests through the backend the server uses (same payload builder, so `LLM_EXTRA_BODY`, the seed and `temperature=0` apply; `max_tokens` cut to at most 128, 15 second timeout). Each asks for one fresh random token (`uuid4` hex); the prompts differ only in that token, so a cache collides on the second. `[PASS]` when each reply contains its own token (extra words around it are fine), with the reply times in the detail and never the reply; `[FAIL]` when the second reply carries the first token or either lacks its own, saying plainly that a response cache returned another prompt's answer, that it can serve the wrong SQL for a similar question, and that the fix is to disable response caching for this client at the gateway or to point `OPENAI_BASE_URL` at the model server directly. `[SKIP]` under `LLM_PROVIDER=mock`, for a remote endpoint while `LLM_ALLOW_REMOTE` is off (nothing is sent), when the endpoint does not answer (`OpenAI-compatible model exists` reports that), and when a reasoning model spends its whole budget before answering. It is a deep check on the admin panel (`GET /admin/health/checks?deep=1`): it is the only check that spends model-server time, and behind a response cache each run would leave two junk entries; the command line always runs it, and the panel's confirmation dialog for the deep checks now says that two small requests go to the model. Runbook §3.1 and §20.10.
- **The setup wizard stops when the model repeats itself, and `--allow-repeated-answers` overrides (PR #181).** Steps 3 (aliases) and 4 (business rules) compare the model's parsed answer for each table (the alias list; the rule text with the table names taken out; empty answers aside). When 3 tables in a row get the identical answer the step stops with `Step 3 stopped: the model endpoint returned the identical alias list for 3 different tables in a row (...)`, naming the suspected response cache, the preflight check and the gateway fix. It writes no file, exits with status 3 and leaves the run resumable (`--resume` continues once the endpoint is fixed). `--allow-repeated-answers` turns the stop into one warning, for a schema so small that identical answers are expected. Each of the two steps ends with a line such as `183 distinct alias lists for 183 tables`. Step 5 asks a single question for the whole schema and has nothing to compare. Runbook §2.2 and the Persian guide.
- **An early warning for a response cache at run time, `LLM_CACHE_SUSPECT_MS` (default `50`, `0` = off) (PR #181).** `observability/response_cache_watch.py` keeps, in memory per process, the last 200 (question hash, SQL hash) pairs per data source. When a different question returns byte-identical SQL within two minutes and the model's reply took less than the threshold, `POST /query` logs one `WARNING` per streak and writes `suspected_cached_response: true` in the audit `llm` block (and the response's). It never blocks or changes a request, since two questions can map to the same SQL; a check is a hash and a dictionary lookup, with an injectable clock for tests. The multi-turn session engine is not watched: its refinements can reuse a previous turn's SQL on purpose. Runbook §20.10.
- **Runbook §20.10, "A response cache in front of the model" (PR #181).** What it is (and how it differs from the model server's prefix KV cache), the symptom (identical answers, replies in milliseconds), the risk (a silent wrong SQL, and an accuracy run that measures the cache), how to confirm, and the fix. Tests: `tests/test_response_cache_detection.py`.
- **`python -m eval.cli run ... --share PATH` writes a share-safe results bundle next to the normal report (`eval/share.py`) (PR #180).** Needs `--live --reference live`; anything else is refused in one line, exit status 2, before the golden file is loaded or the model is called. UTF-8 JSON with sorted keys and cases sorted by id (two rounds `diff` cleanly), written atomically and owner-only. The path must be outside the checkout or under the git-ignored `eval_data/`: every other path in the tracked tree is refused (`check_share_path`, which needs neither git nor a `.gitignore` parser), and `--share` may not be the `--out` file. Header: `format_version`, application version, `prefix_version`, the model name (never the endpoint, host or key; a name that is a URL is cut to its last segment), the golden file's base name and SHA-256 (line endings normalised, so Windows and Linux agree), `PROMPT_RETRIEVAL_TOKEN_BUDGET`, each data source's prompt path (`static` or `retrieval`), start and end time, and a `privacy` block that says what is and is not inside. Per case: id, tags, question, reference and generated SQL, status, expected and selected data source, the guard's rejection reasons (class and redacted message), the correction count, redacted error text, latency and a mismatch diagnosis. A summary: execution accuracy overall, per tag and per data source, source-selection accuracy, the distribution of mismatch classes (every class, zeros included), correction rounds and latency p50/p95. **Never in it:** result rows or cell values, credentials, `.env` values, endpoint URLs, API keys, principal ids, a case's `notes` or recorded rows. **Readable on purpose, and said so in the block:** the question and the result column names (a `PIVOT` turns values into names). Runbook §18.6; Persian guide §۵.۵.
- **`--share-redact-literals` replaces string and numeric literals in both SQL texts (PR #180).** Without it the SQL is verbatim, because diagnosing a wrong filter needs the filter, and the header says so. The literals are found by parsing with sqlglot (`tsql`), never by pattern, so a quote inside a comment or a bracketed name is not mistaken for a string: strings (`N'...'`, hex and bit strings included) become `:str`, numbers `:num`, an `ORDER BY` / `GROUP BY` position is kept, comments are dropped, and text that does not parse is withheld whole rather than guessed at.
- **A mismatch diagnosis, computed on the server from the two result frames and holding no value (`eval/diagnosis.py`, `eval.models.MismatchDiagnosis`) (PR #180).** `mismatch_class` is one of `match`, `row_count_differs`, `column_count_differs`, `column_names_differ`, `same_rows_different_order`, `values_differ_same_shape`, `actual_is_subset`, `actual_is_superset`, `actual_empty`, `expected_empty` or `not_executed`, with the row counts, the column names of both results (cut at 128 characters and 64 names) and, for two single numeric cells that differ, `relative_error_bucket` (`<1%`, `<10%`, `>=10%`; never the numbers). The decision order is in the module docstring; it uses the same tolerance and ordering rule as the verdict (`eval.compare`), whose new `multiset_equal`, `multiset_contains` and `align_columns` tell "the same rows in another order" and "the same columns in another order" from "other values". It is also in every result of the normal `--out` report (`diagnosis`, `corrections`, `guard_rejections`; a report or baseline written before loads as before).
- **`python -m eval.cli compare OLD NEW` compares two rounds by `case_id` (PR #180).** Reads share bundles or `--out` reports and prints ids and counts only (no question, SQL or value): fixed, regressed, still failing, added and removed cases; the accuracy delta overall, on the cases both rounds have and per tag; the shift in each mismatch class; and a note when the golden file or the `--reference` differs. `--split dev|test` with `--split-file` (one id per line, a JSON list, or a JSON object with `dev` and `test` lists) limits it to one half; `--fail-on-regression` exits `1` when a compared case passed before and fails now, so `--split test --fail-on-regression` gates the held-out half; `--json` prints the comparison as JSON. Exit status `2` for an unreadable input. It needs no model, database or `project_config/`.
- **`python -m eval.cli split --golden FILE --out DIR` writes a deterministic dev / test split (`eval/split.py`) (PR #180).** Within each `tags[0]` category, ids sorted, even positions to dev and odd to test; nothing is random and the order of the file does not matter. Writes `dev_ids.txt`, `test_ids.txt` and `split.json` (counts per category, the golden file's name and hash) and prints the counts. Every case is placed whatever its `status`, so activating cases later moves nobody. Re-running on the same set is a no-op; on a changed set it refuses until `--force`, because the held-out half must not move between rounds.
- **`core.redaction.redact_error_text` and `redact_sql_literals` (PR #180).** The one redactor for text that leaves the server beside `redact_db_url`: quoted strings become `'…'` (an unterminated one to the end of the text), runs of four or more digits (Persian and Arabic-Indic too) become `#`, `PWD=` values and `user:password@` are masked first, the result is cut at 600 characters; and the SQL-literal redaction above.
- **`eval.cli run --live` generates through the server's own pipeline by default (`eval/pipeline.py`); `--single-shot` keeps the one-call behaviour (PR #180).** The default runs `llm.sql_agent.SQLAgent` (as `api.runner` builds it, with the router from the `LLM_*` settings): the guard, correctable and policy rejections (a policy rejection is not retried), up to `MAX_CORRECTION_ATTEMPTS` correction rounds with the 6.11.0 correction hints, source routing, the row cap the agent injects into a statement without one (`DEFAULT_TOP_N`), and `database.executor.execute_sql` with its query timeout, rolled-back transaction and `MAX_ROWS_RETURNED`. A round is a statement that the guard rejected or that failed to run. `TracingSQLAgent` overrides only the guard step to count rounds and keep each rejection; the generator hands back the result the agent already obtained for the final SQL (so it is not executed twice) and the reference query still goes to the real executor. A run that ran out of rounds is `guard_rejected` or `execution_error` with the last candidate SQL; a transport failure stays `generation_error`. **No principal policy:** the evaluation passes no `denied_columns`, so a deployment whose keys deny columns is measured as if they did not. `--single-shot` is the old behaviour, a model-only number; `--structured` needs it, and `--single-shot` needs `--live`. The mode is `pipeline` (`production` or `single_shot`) in the `--out` report and `run.pipeline` in the share file, and printed in the report; offline runs have none and are unchanged (no CI job runs `--live`).
- **A result cut by the row cap passes as `match_capped` (`eval/diagnosis.py`) (PR #180).** A result with exactly the cap's row count (the `TOP n` the agent injected, or `MAX_ROWS_RETURNED` if lower), fewer rows than the reference and every row a row of the reference under `compare`'s matching rules (for a reference with `ORDER BY ... TOP`, its first rows in that order) is scored `pass`: the analyst saw the right rows, truncated by policy. The diagnosis records `capped: true` and `row_cap` (a count; no row value is exported), and the class is in the summary distribution (12 classes now). **A count confirms it before it passes (`eval/capcheck.py`):** one more read-only query, `SELECT COUNT(*) FROM (<the model's SQL before the cap was injected>) AS q`, built with sqlglot (`tsql`; an `ORDER BY` without `TOP`/`OFFSET` is removed, a `WITH` clause hoisted and repeated or unnamed columns named, as SQL Server requires of a derived table) and run through the same executor, so the timeout and the read-only transaction apply. The case passes only if the count equals the reference's row count (`uncapped_rows`); an over-filtered query that returns a cap's worth of valid rows has a smaller count and fails as `actual_is_subset`. A count that cannot be built or run fails the case as `actual_is_subset` with `capped_unverified: true`, never a silent pass; when the reference result itself reached `MAX_ROWS_RETURNED` (so its length is not its size) the reference is counted the same way, `SELECT COUNT(*) FROM (<reference SQL>) AS q`, and the two counts must be equal (`uncapped_rows == reference_uncapped_rows`, also in the diagnosis; no "at least" rule), with a reference count that cannot run failing as `capped_unverified`. A row not in the reference, or a subset short of the cap, is still a failure (`actual_is_subset` and the rest); `--single-shot` has no cap, so the class never appears there. Tests: `eval/tests/test_diagnosis.py` (`TestMatchCapped`), `eval/tests/test_capcheck.py` (including a cut reference with equal, different and failing counts), `eval/tests/test_pipeline.py`.
- **`corrections`, `guard_rejections` and `execution_errors` per case are real, round by round (PR #180).** `guard_rejections` is `[{round, class, message}]` (round 1 is the first try) and `execution_errors` the same for statements that passed the guard and failed to run, so a repeated rejection is visible. In the share file the messages are redacted like the error text; the summary has `correction_rounds`, `guard_rejections` and `execution_errors`. `--single-shot` reports `corrections` 0 by construction, and its bundle note says so.
- **`compare` refuses two rounds whose `run.pipeline` differs; `--allow-mixed` compares them with a note (PR #180).** Reports written by `--out` carry the same field. A baseline (`--baseline`) is refused across pipelines too, like `--reference`; one written before the field existed counts as `single_shot`.
- **`eval.cli run --share ... --share-only-ids FILE` keeps the held-out half's content on the server (PR #180).** `FILE` is a list of case ids (`dev_ids.txt` from `split`, one id per line, or a JSON list). Only those cases carry the question, SQL, error text and diagnosis; every other case carries only `case_id`, `tags`, `passed` and `mismatch_class`, and the summary is computed from all cases first. The header has `detail` (`scope`, `cases_with_detail`, `cases_total`) and the `privacy` block says so; `compare --split test` works on such a file. An id not in the golden file is a warning; an empty or missing file is exit status 2. Needs `--share`.
- Documentation (PR #180): runbook §18.6 "Accuracy rounds with a reviewer" (the commands in order, what to send and what never to, how fixes come back, `compare`, the stop criteria and a privacy paragraph) with pointers in §18 and §19; the same in Persian, §۵.۵ of `docs/fa/getting-started.md`. Tests: `eval/tests/test_pipeline.py` (the correction loop with a fake model that fails once, repeated and exhausted rounds, the row cap, `--single-shot` making one call, the CLI modes, baseline across pipelines), `eval/tests/test_diagnosis.py` (every class), `eval/tests/test_share.py` (a whole `--live --reference live` run on SQLite with a fake model: no sentinel cell value in the file, with and without `--share-redact-literals`; error redaction; the refusals; the path guard; file mode and stable bytes), `eval/tests/test_rounds.py` (`compare` and `split`), `eval/tests/test_cli_startup.py`, `tests/test_redaction.py`.

### Changed

- **`eval.cli run --live` no longer measures a single model call (PR #180).** It used to generate once and apply the guard once, a model-only figure below what analysts get, because the server corrects a rejected or failing statement up to `MAX_CORRECTION_ATTEMPTS` times and caps rows. The default is now the server's pipeline (see Added); the old behaviour is `--single-shot`. `--live` figures recorded before this change are single-shot figures. `--structured` now requires `--single-shot` with `--live`.

### Fixed

- **A setup run against a model behind a response cache no longer writes one answer for every table (PR #181).** An operator ran the wizard against two databases through a gateway that kept a response cache: most tables of the first got the byte-identical alias list the model had given for the first table, the second database got that list again, and the business rules read as nonsense (`Value column: X. Volume column: X.` on a table where that makes no sense). The wizard wrote all of it without a word, because nothing in the reply says it came from a cache. Steps 3 and 4 now stop at the third identical answer in a row (above).
- **`python -m eval.cli --help` died in an import-time traceback on a machine with no, or a broken, `project_config/` (PR #180).** `eval.cli` imported the runner at module scope, which imports the prompt builder, which binds `knowledge.business_rules`. Everything that reads `project_config/` is now imported inside the command that needs it, so `--help`, `split` and `compare` need none. `run`, `verify` and `recall` check the files first (`core.config_preflight.exit_if_project_config_unusable`, as `scripts/prompt_budget.py` and `api/startup_preflight.py` do) and print one `cannot load project_config: ...` line with exit status 2. To make that possible `load_golden_cases` moved to `eval.store` (still importable from `eval.runner`) and `eval.determinism` imports the runner for type checking only. Tests: `eval/tests/test_cli_startup.py` (fresh interpreters: no `Traceback` in stderr for `--help`, nor, for a broken config, a second line).

### Upgrading

- **The live evaluation now measures the production pipeline by default (PR #180).** `python -m eval.cli run --live` generates through the server's own loop (guard, correction rounds, row cap), so its accuracy is the analyst's and is higher than a single-shot figure for the same model. A figure, a `--save-baseline` file or a share bundle recorded before this release is single-shot: `--baseline` and `compare` refuse to set it against a production run (`compare --allow-mixed` overrides, with a note). Record a new baseline on 6.12.0 after the restart, in the mode you will gate on (`--live --reference live --save-baseline ...`); add `--single-shot` to every command that must stay model-only, which still compares with a baseline recorded before the upgrade. Offline runs (CI) are unchanged. The run uses no per-key `denied_columns` policy, so a deployment that denies columns is measured as if it did not. For the rounds with a reviewer, add `--share-only-ids eval_data/split/dev_ids.txt` so the test half's content never leaves the server (runbook §18.6).
- **Run `python scripts/verify_deployment.py` after the upgrade, before trusting answers or running an accuracy round (PR #181).** The new check `Model endpoint answers each prompt itself` sends two tiny requests to the model through the backend the server uses. If it fails, a response cache in front of the model returned another prompt's answer: fix the gateway first (disable response caching for this client, or point `OPENAI_BASE_URL` and `WIZARD_LLM_BASE_URL` at the model server directly; runbook §20.10) and run the command again until it passes. It is skipped under `LLM_PROVIDER=mock` and, for a remote endpoint, while `LLM_ALLOW_REMOTE` is off; on the admin panel it is a deep check. The setup wizard also stops with `Step 3 stopped` when the model repeats itself (`--allow-repeated-answers` overrides, for a schema so small that identical answers are expected); a resumed run (`--resume`) continues once the endpoint is fixed.
- **One setting is new and optional: `LLM_CACHE_SUSPECT_MS` (default `50`, `0` = off) (PR #181).** It adds `suspected_cached_response: true` to the audit `llm` block (and the response's) of a request that returned byte-identical SQL to a different question faster than the threshold, and one `WARNING` per streak; it never blocks or changes a request. A reader of `audit_log.jsonl` that ignores unknown keys is unaffected, and records written earlier simply lack it. Set `0` to turn the warning off. It is in `.env.example`; no existing setting was renamed and no default changed.
- **No `requirements.lock` pin changed (PR #180, PR #181).** `git diff v6.11.0 -- requirements.lock` is empty. Run `pip install -r requirements.lock` as for every upgrade, but nothing new is installed, and no file format changed: a `project_config/` that loads today loads the same way.

## [6.11.0] — 2026-10-14

Table selection is now measured on a real, operator-written question set, and fixed. The largest cause of misses was that the retrieval configuration named tables by their bare name where `schema.yaml` keys them `Schema.Name`; configured names are now resolved to the schema-qualified keys, with smaller fixes for inflected and reordered aliases and for a year that never says the word. Recall on the held-out half of the question set went from 0.62 to 0.98; the questions and the tables stay with the deployment that measured them. Five new preflight checks tell a raw `database.schema_inspector_cli` draft from a curated `project_config/` (a draft copied over the curated files loads, starts and answers worse, and nothing used to say why); three can fail and two only advise. On the model side, `LLM_PREFIX_KEEPWARM_SECONDS` periodically re-sends the static prefix so a shared model server does not evict its cache between questions, the static prefix is about 2% smaller by formatting alone, `prefix_cache_hit` follows the `cached_tokens` the server reports instead of a prompt-size guess, and a correction round is given the real column names of the table the guard rejected.

### Added

- **`LLM_PREFIX_KEEPWARM_SECONDS` keeps the model server's prefix cache warm between questions (PR #178).** Default `0` (off). Above zero, a background thread (`llm.warmup.KeepWarm`) re-sends each static-prefix source's prefix every that many seconds: the request the start-up warm-up sends, built by the same code (`warm_prefix_cache`), so it cannot drift from it. **A source a real question used within the last interval is skipped** (`llm/prefix_activity.py` records, per prefix hash and never its text, when the router last served one), so under traffic the keep-warm sends nothing. It follows every rule of the start-up warm-up: nothing for a retrieval-path source, nothing under `LLM_PROVIDER=mock`, nothing to an untrusted endpoint without `LLM_ALLOW_REMOTE`, a failure is a log line and never fatal. A repeated outcome is logged once and then at `DEBUG` until it changes, so an endpoint that stays down is one warning and the pass that succeeds again says how many passes failed. It pauses while maintenance mode is on and stops with the server. `Settings.validate()` refuses a value that is neither `0` nor at least `30`. `300` is recommended on a shared model server; runbook §20.7 gives the reasoning, the cost under traffic (none) and idle (about 12 small requests an hour per source), and how to check from `cached_tokens` that the interval is short enough.
- **`GET /admin/llm/warmup` shows the keep-warm's last run and outcome; `POST /admin/llm/warmup` returns the same under `keepwarm` (PR #178).** `operations` or `security` keys can read it: `enabled` (with a `reason` when it is not), `interval_seconds`, `running`, `paused`, the number of `runs`, `last_run_at` (UTC), `last_run_outcome` (`warmed`, `skipped`, `failed`, `paused` or `busy`) and one entry per data source in `last_run`, the entries `POST` already returned. No prompt text, endpoint or key.
- **Runbook §20.8 says what `prefix_cache_hit` is and is not (PR #178).** The audit field is `prompt_tokens < 0.5 × the prefix estimate`, a rule made for a backend that reported only the tokens it computed. An OpenAI-compatible server reports the whole prompt whether or not it came from the cache, so today `true` means a short prompt and `false` a prompt about as large as the whole prefix, cached or not. The section gives the reading that works (the request's time against its size, and `cached_tokens` where the server reports it) and corrects the §20.4 row that took `false` for a cold cache.
- **The preflight now says when `project_config/` is a raw draft of `database.schema_inspector_cli` instead of the curated configuration (PR #177).** An operator who copied the inspector's drafts over a curated `project_config/` got a deployment that loaded (6.9.3 made the drafts valid YAML), started and answered worse, and nothing said why: 75 curated tables in two databases had become 278 tables from one, about 200 of them staging and change-capture tables; one database was gone; every description was the placeholder `<schema.table> — auto-detected <kind> table (~N rows)`; column notes quoted live sample values (phone numbers, street addresses); `aliases.yaml` had an empty `ring_aliases` and synonym seeds taken from sample values. Five checks in `scripts/verify_deployment.py` (`core/config_doctor.py` reads the files), after `project_config/ loads` and before `Rate limit sane for deployment`; each reads files only, with no database call, so the admin health card runs them without "deep checks":
  - `project_config/ is curated, not a raw draft` fails when any `*.yaml` file of `project_config/` still carries the `# AUTO-GENERATED DRAFT` header (the first ten lines are searched), or when more than a quarter of `schema.yaml`'s table descriptions are the inspector's placeholder. The detail names the files and the counts, says to restore the curated files and add tables with `python scripts/sync_schema.py --add-tables 'Schema.*'`, and points to the runbook (§2.3, §16.3); a file that was really reviewed passes once its header line is deleted. A table that `--add-tables` appended is described by its own name, so a curated file that grew that way stays curated. `schema.yaml` is read even when a repeated key (a draft from before 6.9.3) keeps the application from loading it, every occurrence counted.
  - `Aliases are curated` fails when more than half of `aliases.yaml`'s synonym keys are not words (no run of two letters: timestamps, numbers), or when more than half of its synonym lines still end in the inspector's `# REVIEW` marker (a seed is a sample value of any text column, so only part of a real draft is timestamps; all of it carries the marker). An `aliases.yaml` that is not YAML (a quote in a sample value, before 6.9.3) is read from its lines. An empty `ring_aliases` is advice, not a failure: only `ValueRetriever.extract_ring` reads it, to set the `Ring` filter from a ring (hall) name in the question, so a warehouse without that dimension leaves it empty, as the setup wizard does when it finds no dimension table; the `[PASS]` line says `advice: ring_aliases is empty` while `schema.yaml` or `entities.yaml` exist.
  - `Every data source is described` fails when a source of `datasources.yaml` (or the one `default` source without the file) has no table in `schema.yaml`, a table belonging to the source its `datasource:` line names (to each, for a list) and to the default source when it has none. That is a whole database lost; the detail names the source and the way back.
  - `No staging or change-capture tables` and `No live sample values in descriptions` are advice: they never fail. There is no `WARN` status; advice is a `[PASS]` line whose detail starts with `advice:`, which leaves the exit code, the CLI summary and the admin card (it shows any detail, and has no other status to show) as they are. The first counts the tables per schema that match `STAGING_PATTERNS`; the second counts, as phone-like (7 or more digits, Persian digits too), national-id-like (exactly 10) and address-like (a long segment of words with digits), the notes whose `e.g.` list quotes a value (an `e.g.` inside parentheses is prose, and dates are left alone) and lists the table keys. Neither detail ever contains a value from a file.
- **`STAGING_PATTERNS` sets what counts as a staging or change-capture table (PR #177).** Comma-separated `fnmatch` patterns, case-insensitive; a bare pattern is matched against the table's schema, a `table:` prefix against its name. The default is `*_Stg,*stg*,*staging*,table:*_Inserts,table:*_Updates,table:*_Deletes`; an empty value turns `No staging or change-capture tables` off. In `.env.example` and `config.Settings.staging_patterns`.
- Documentation (PR #177): `docs/deployment-runbook.md` §3 and §3.1 describe the five checks and the advice convention, and §17 step 3 lists the new refusal with the other things the preflight catches after an upgrade; the Persian guide (`docs/fa/getting-started.md`, §۰.۴ and §۴) says the same in Persian. Tests: `tests/test_config_doctor.py` (thresholds, empty and unreadable files, a single source, the patterns setting, no value in any message), `tests/test_setup_on_warehouse_shape.py` (`TestPreflightOnAnInspectorDraft`: the inspector's real drafts of the warehouse-shaped fixture make every check fire, and `TestPreflightOnTheCuratedConfig`: its curated config passes), `tests/test_admin.py` (`TestCurationChecksOnTheCard`).

### Changed

- **The static prefix is smaller by formatting alone (about 2% of a large schema) (PR #178).** A column line is `- Name: description` instead of `  - Name: description`, and each business rule is stripped before the rules are joined, so a rule's trailing newline no longer adds a blank line between rules. Nothing the model reads is removed or reworded; `tests/test_prefix_compaction.py` checks that every table, description, column, relationship, rule, metric and example is still there, and reverses the two changes to get the previous release's prefix back byte for byte. The retrieval path's schema block is the same function, so it is that much smaller too. On the synthetic `project_config.example` the prefix estimate goes from 1,815 to 1,798 tokens (7,260 to 7,193 characters); the saving is per column, so a schema with hundreds of columns saves hundreds of tokens. The golden hashes in `tests/test_source_prompts.py` were re-captured on purpose, for this reason.
- **Table names in `entities.yaml`, `fact_tables`, `fact_patterns` and `always_include` are resolved to `schema.yaml` keys (`retrieval/table_keys.py`) (PR #176).** A bare name resolves the way the SQL guard resolves a reference in a query: an exact key (brackets and case ignored) is itself; otherwise the keys with that bare name, narrowed by any qualifier written (`Dim.Customer`). A name that matches several keys is read in the schemas the configuration's other, unambiguous names live in, or in the data source whose `keywords:` the question contains (`datasources.yaml`); when neither narrows it, every candidate is kept; a name that matches no key is kept as written. `ContextRetriever.entities` and `.facts`, `RetrievalContext.selected_tables`, the pruning stage, join expansion's choice of the fact table and the fact/dimension split of the extras and of the TF-IDF fallback all use the keys. A configuration that already writes full keys is unchanged. Tests: `tests/test_table_keys.py`.
- **Alias matching also compares whole words after plural stemming, and a multi-word alias matches its words in any order (`retrieval/table_evidence.py`) (PR #176).** The substring test stays; beside it `city` finds `cities` and the plural alias `cities` finds "city" (`-ies` for `-y`, and a plural alias for a singular question), and an alias such as `gear tooth pitch` is found in "pitch of each tooth on the gear" because all its stemmed words occur. A different word is not an inflection (`order` is not found in `ordinary`). Tests: `tests/test_table_evidence.py` (`TestAliasHits`).
- **A four-digit calendar year in a question adds the words `year` and `سال` to the expanded question (`schema_data/retriever.py`) (PR #176).** "Totals in 2023" is about a period but never says the word, so the date dimension's `always_include` triggers and the descriptions did not see it. The years are 1300-1599 (Solar Hijri) and 1900-2199 (Gregorian), Persian and Arabic-Indic digits included; a number such as `2000` in "orders above 2000" is also taken for a year, at the cost of one extra table. Tests: `tests/test_retriever.py` (`TestExpandSynonyms`).
- **A correction prompt names what exists when the guard rejected a name that does not (`llm/correction_hints.py`) (PR #178).** In the real-use audit log a request that began with an unknown-column or unknown-table rejection used both correction rounds (three model calls) and was refused in the end every time; the retry carried only the guard's message, so the model guessed again. The retry now adds, deterministically and from `schema.yaml`: for an unknown column, the table's real column names (names only, in `schema.yaml` order, `... and N more` past 200); for an unknown table, the closest real tables as the reference to write in `FROM` (limited to the question's data source). A column the key's `denied_columns` fully denies is left out; a join-only column is kept and marked `(join key only)`, the one use the guard allows; an unreadable policy adds nothing. The text goes to the question segment of the retry like the rest of the correction text, never the static prefix, and is not repeated once shown. `CorrectableRejection` for an unknown column now carries the `table` it is about (`SqlGuardRejection.table`). Every other error gets the correction text it got before, byte for byte. Both loops (`SQLAgent`, `session/engine.py`) use it. Runbook §20.9.
- **`prefix_cache_hit` in the audit `llm` block follows the model server's `cached_tokens` when it reports one (PR #178).** `usage.prompt_tokens_details.cached_tokens` (vLLM with `--enable-prompt-tokens-details`) is recorded as `cached_tokens`, `prefix_cache_hit` is true when it covers at least half of the static prefix (of the whole prompt when shorter), and `prefix_cache_hit_source` says `"server"`. Without it the older prompt-size rule is the fallback and the source is `"estimate"`; a record written before these fields reads as an estimate. That rule compared `prompt_tokens` with the prefix estimate, which was a cache signal for a backend that reported only the tokens it computed and, on an OpenAI-compatible server that counts the whole prompt, says only that the prompt was short: in the real-use log a full-size prompt served from the cache in under a second counted as a miss and short prompts that were computed from scratch counted as hits. `scripts/analyze_audit_log.py` prints a real hit rate over the calls that reported it and the `llm` stage's time for `hit`, `miss` and `unreported` calls (`latency_by_cache`, also `cache_behaviour.server_hit_rate` and the quality report's `latency.llm_by_cache`); the admin prefix-reuse tile prefers the server rate and names its source, and the web badge marks an estimate `(est.)`. `docs/api-contract-v2.md` §6 and runbook §20.8.

### Fixed

- **A retrieval configuration that names tables by their bare name no longer loses the tables it names when `schema.yaml` qualifies the keys (PR #176).** A warehouse that repeats table names across schemas has keys such as `sales.Customer`; an `entities.yaml` written before that still said `Customer`. The alias retriever returned `Customer`, which is not a key: the schema block skips an unknown name, the ranking never scored it, the pruning stage kept it for the wrong reason, and `name in FACT_TABLES` was false for every key, so the extras and the TF-IDF fallback could not tell facts from dimensions. A plain count of one dimension table retrieved four other tables and not that one. It was by far the largest cause of the misses on the real question set that found it (`docs/design/RETRIEVAL.md` §8 has the method and the generic mechanisms; the figures stay with that deployment). Tests: `tests/test_table_keys.py`.
- **`ContextRetriever` finds an entity's label and its static value filter through the table key (PR #176).** The «list of values unavailable» warning looked the label up by the table key in `ENTITIES`, and the dimension-vocabulary pass was skipped for a table only when `ValueRetriever` had keyed a filter by that exact table name; both now go through the same resolution, so a bare entity name still gives its label and still makes the static filter win. Tests: `tests/test_table_keys.py` (`TestContextRetrieverWithBareEntityNames`).

### Upgrading

- **The new preflight checks may fail on a `project_config/` that has no `datasource:` lines when there is more than one data source (PR #177).** `Every data source is described` counts a table with no `datasource:` line toward the default source only, so a `datasources.yaml` that lists a second source while `schema.yaml` has no such line reports the second source as having no table. Run `python scripts/sync_schema.py`, which writes the lines; review `schema.synced.yaml`, replace `schema.yaml` and run `python scripts/verify_deployment.py` again (runbook §16.3). A raw `database.schema_inspector_cli` draft copied over the curated files fails `project_config/ is curated, not a raw draft` as well: restore the curated files, then add tables with `python scripts/sync_schema.py --add-tables 'Schema.*'`. The other checks only advise.
- **The prefix text changes once, so `prefix_version` and every cached answer's key change once (PR #178).** Answers cached by the previous release are not found after the restart, and the model server's prefix cache is cold for the new text. Run `POST /admin/llm/warmup` after restarting (or leave `LLM_PREFIX_WARMUP_ON_STARTUP` on, which does it). Nothing else needs doing; no file format changed.
- **A source whose prefix estimate was within about 2% above `PROMPT_RETRIEVAL_TOKEN_BUDGET` may now take the static path instead of retrieval (PR #178).** The estimate dropped by the saving above. Run `python scripts/prompt_budget.py` if a source sits that close to the budget; the line it prints says which path each source takes.
- **Add `--enable-prompt-tokens-details` to the vLLM launch command if you want `prefix_cache_hit` to be truthful (optional) (PR #178).** With it the server reports `cached_tokens` and `prefix_cache_hit` is the server's own answer. Without it `cached_tokens` stays `null`, `prefix_cache_hit_source` stays `"estimate"` and the analysis reports the real rate as `n/a`. Two new keys appear in the audit `llm` block (`prefix_cache_hit_source`, `cached_tokens`); a reader of the log that ignores unknown keys is unaffected, and records written earlier simply lack them.
- **The text of a correction round changes for an unknown column or table (and only then) (PR #178).** Nothing to configure; the retry is a few dozen tokens longer for those two rejections.
- **Set `LLM_PREFIX_KEEPWARM_SECONDS=300` on a model server that other people also use (optional) (PR #178).** Left at `0` nothing is sent, as before; the value must be `0` or at least `30`. Leave it at `0` when this deployment is the only user of the model server, when the server does not cache prefixes, or when every source is on the retrieval path.
- **Retrieval can select different tables by default for a configuration that names tables by their bare name (recall is the point; precision and token use follow it) (PR #176).** Run `python -m eval.cli recall --golden eval_data/golden.jsonl` before and after, as in the runbook (§18.4); on such a configuration the tables the aliases name are now in the prompt, so the mean number of retrieved tables rises. A configuration that writes full `Schema.Name` keys everywhere selects what it did before, apart from the inflection, word-order and year rules above.
- **The dimension vocabulary is now consulted for entities whose bare name used to resolve to nothing (PR #176).** `PREFETCH_COLUMNS` is keyed by schema key, so for `entities.yaml` written with bare names the prefetch and the match against the question's words were silently skipped; they now run for the tables that declare `prefetchable_columns`, as they always did for a configuration with full keys. Nothing is read from the warehouse on the question path (the cache is filled in the background, as before).
- **If a bare name in `entities.yaml` or `retrieval_hints.yaml` matches tables in several schemas, qualify it (`sales.Customer`) when the schema the other names live in is not the one you mean (PR #176).** The rule above reads it in the schemas of the file's other names, and keeps every candidate when it cannot tell.
- **Two settings are new, both optional, and no existing setting changed (PR #177, PR #178).** `STAGING_PATTERNS` (advice only) and `LLM_PREFIX_KEEPWARM_SECONDS` (default `0`, off) are added to `.env.example`; no other setting is renamed and no default changed (the `RETRIEVAL_*` settings keep theirs, `docs/design/RETRIEVAL.md` §6), and no `requirements.lock` pin changed. A `project_config/` that loads today loads the same way, apart from the preflight note above.

## [6.10.0] — 2026-10-13

Analysts can now confirm a correct answer with one click, and each confirmation becomes a reviewed case in the golden set, so everyday use builds the accuracy gate's reference instead of a separate review session. A real-use quality report, in `scripts/analyze_audit_log.py` and as a new admin card, joins the audit log with the confirmations and flags: execution success, guard-rejection, correction and empty-result rates, the confirmed share of rated answers, and latency per stage. Every setup tool is now tested on a warehouse-shaped fixture, which found and fixed three bugs: the inspector's entities, aliases and relationships drafts broke on a sample value with a quote or a backslash, an empty `relationships.yaml` crashed join-path building, and `database.catalogue` showed none of the tables in an attached SQLite database.

### Added

- **"This answer is correct" (پاسخ درست است): analysts confirm a correct answer with one click (PR #174).** The golden set stays empty because building it costs analysts a separate review session; everyday use now builds it. `POST` and `DELETE /v2/sessions/{sid}/turns/{tid}/confirmation` (and `GET /v2/sessions/{sid}/confirmations`, so a reopened transcript shows each turn's state) store a confirmation in the new `turn_confirmations` table of the application database (Alembic revision `0005_turn_confirmations`; created at start-up like the other tables; carried by `scripts/migrate_app_db.py`). Like a wrong-answer flag it holds identifiers only (`session_id`, `turn_id`, `principal_id`, `request_id` resolved from the audit record, `created_at`), never the question, the SQL or any data; there is at most one per turn and principal, confirming again is idempotent, and a principal withdraws only their own. Only the turn's owner may confirm it (404 for anyone else, as for flagging), and only a turn whose audit record shows SQL that ran and returned a result (`409` otherwise). A principal who flags a turn they had confirmed, or confirms one they had flagged, keeps only the later action: the earlier one is removed (an admin-resolved flag is kept, since triage recorded a decision on it). The web UI shows a «پاسخ درست است» toggle (`aria-pressed`, keyboard reachable) on the answer itself, in a footer under the result table or chart and outside the collapsed details drawer, on successful answers only; the wrong-answer flag stays in the drawer. The toggle keeps its box while it changes state, and its confirmed state (a check mark, a thicker border and a tint) shows in both themes. `docs/api-contract-v2.md` §3 documents the routes; `docs/deployment-runbook.md` §18.1.1 and the Persian guide §۵.۱.۱ show the flow. Tests: `tests/test_appdb_confirmations.py`, `tests/test_v2_confirmation_routes.py`, `tests/web_ui/test_web_ui_confirmation_control.py`, `tests/web_ui/test_web_ui_confirmation_placement.py`.
- **`scripts/harvest_golden.py` harvests confirmed answers as reviewed golden cases (PR #174).** On by default (`--no-confirmations` opts out). Each confirmed turn becomes a case with `status = "reviewed"` (a person confirmed its `expected_sql`, the SQL that ran; `eval.cli verify` has not run it yet), the confirming principal's id and the confirmation time in its `notes`, and the tag `origin:confirmed`. Confirmed cases come before the stratified sample and are never dropped by `--n`; two confirmations of the same normalised question are one case, the most recent one; a confirmation whose audit record has rotated away is counted and skipped. `scripts/golden_sheet.py export` lists them with the verdict already `correct`, and `python -m eval.cli verify --accept` is still the gate that activates them. The console output stays counts only, and the file stays owner-only in `eval_data/`. Tests: `tests/test_confirmed_golden_cases.py`.
- **A real-use quality report, in `scripts/analyze_audit_log.py` and in the admin panel (PR #174).** `observability.quality.build_quality_report` joins the audit log with the confirmations and flags over the last 7 days (`--quality-days N`) and all time, overall and per data source: execution success, guard-rejection, correction, empty-result and out-of-scope rates, error rate by class (reusing `failure_taxonomy`), confirmed and flagged answers with the confirmed share of rated answers (labelled a lower-bound signal, not accuracy), and end-to-end and per-stage latency p50/p95 including the `llm` split into `ttft_ms` and `generation_ms` (reusing `latency_report` and `llm_latency_split`; the split needs `LLM_STREAM_TIMINGS=true`). It copes with an empty log and with records from older versions that lack fields. `analyze_audit_log.py` prints it as a new section (and under `quality` with `--json`), counts and percentiles only. The read-only `GET /admin/quality?days=7` (operations or security, like `/admin/usage`) serves it to a new admin card «کیفیت در استفاده‌ی واقعی», which notes that the confirmed share is not accuracy and that the stage split appears only with `LLM_STREAM_TIMINGS=true`. The route reads the whole audit log, so it is cached per window for `ADMIN_EXPENSIVE_CACHE_TTL_SECONDS` and the card stays in the 30-second auto-refresh instead of joining `EXPENSIVE_CARDS`; its refresh button recomputes. `docs/deployment-runbook.md` §20.6, the Persian guide §۷ and `docs/admin-panel-architecture.md` §3 describe it. Tests: `tests/test_observability_quality.py`, `tests/test_admin_quality_route.py`, `tests/web_ui/test_web_ui_admin_quality_card.py`.

### Changed

- **Every setup tool is now tested on a warehouse-shaped fixture (PR #173).** The suite's fixtures were small, tidy sample schemas, and twice the first real use of a setup tool found a bug they could not. `tests/_warehouse_shape.py` builds, once per session and deterministically, two SQLite data sources with the shape of a real warehouse and invented names: 262 and 75 tables in ten and seven schemas (fact, dimension and staging schemas, change-capture tables, `Replay`, `Alerts`, `Metadata`, `dbo`), the same name in up to four schemas, one dimension in both sources, no declared foreign keys, empty tables, Persian text, and a name with a space and one with a dot; and a curated 34-table `project_config/`. `tests/test_setup_on_warehouse_shape.py` runs `database.schema_inspector_cli`, `scripts/sync_schema.py`, `scripts/assign_datasources.py`, `scripts/prompt_budget.py`, the `scripts/verify_deployment.py` checks that need no SQL Server and `python -m eval.cli recall` on it. They must stay green; the README's Tests section and the tutorial's "Writing tests" say so.

### Fixed

- **The entities, aliases and relationships drafts of `database.schema_inspector_cli` no longer break on a sample value with a quote or a backslash (PR #173).** Only the `schema.yaml` draft wrote its values through `_yaml_str`; the other three wrote `"{value}"` by hand, so the first value read from a live row such as `Pipe 5" steel` or `Path\to` made `entities.yaml` (its `sample_values`) or `aliases.yaml` (a synonym key) unreadable, and a table, schema or column name with a quote did the same in `relationships.yaml`. Every table, schema, column, type, sample value and reference is now quoted with the same helper as `schema.yaml`. Tests: `tests/test_schema_inspector.py` (`TestDraftsQuoteTheValuesTheyQuote`).
- **A `relationships.yaml` with no entry no longer crashes join-path building (PR #173).** The inspector's draft for a database that declares no foreign key was a bare `relationships:`, which is YAML null, and `retrieval.join_paths` (which builds the join graph for questions) and `database.relationship_map` iterated it as a list and raised `TypeError`; a hand-edited file whose last entry was commented out did the same. The draft now writes `relationships: []`, and both readers treat a null like an empty list. Tests: `tests/test_schema_inspector.py`, `tests/test_join_paths.py`, `tests/test_relationship_map.py`.
- **`database.catalogue` reads every `ATTACH`ed SQLite database as a schema (PR #173).** `read_catalogue`, `list_tables` and `list_columns` reflected only the connection's own database on a dialect other than SQL Server, so a SQLite source with attached schemas (the tests' stand-in for a warehouse with many schemas) showed none of its tables: `sync_schema.py` called every one absent and the "table lives in another data source" hint of `Tables are in their data source` found nothing. Each attached database is now a schema of its own name; the connection's own database is still schema `""`, and no other dialect changes. SQL Server is unaffected. Tests: `tests/test_catalogue.py` (`TestAttachedSqliteDatabasesAreSchemas`).

### Upgrading

- **The new `turn_confirmations` table is created automatically when the application starts (PR #174),** like the other application-database tables, so the default setup needs no manual step.
- **Deployments that manage the application database schema with Alembic instead should run `alembic upgrade head` from the repository root (PR #174);** it applies migration `0005_turn_confirmations`.
- **Tell analysts about the «پاسخ درست است» toggle, and review what they confirm (PR #174).** Any analyst can confirm their own answers; there is no separate permission for it. A confirmed answer becomes a reviewed golden case (`status = "reviewed"`, tag `origin:confirmed`), and `python -m eval.cli verify --accept` checks that its SQL runs, not that the SQL is right, so a wrong confirmation would become part of the accuracy gate. Check the `origin:confirmed` rows of `scripts/golden_sheet.py export` before accepting them, or harvest with `--no-confirmations` until you trust the confirmations you see. `docs/deployment-runbook.md` §18.1.1 shows the flow.
- **Set `LLM_STREAM_TIMINGS=true` if the quality card should show the time-to-first-token and generation split (PR #174).** Without it the card still shows end-to-end and per-stage latency, and notes that the split appears only with that setting.
- **Nothing else changes (PR #173, PR #174).** No setting is renamed and no `requirements.lock` pin changed. A `project_config/` that loads today loads the same way.

## [6.9.3] — 2026-10-13

A `schema.yaml` drafted by `database.schema_inspector_cli` is now valid for a warehouse that repeats table names across schemas (every table carries `db_schema`, shared names are keyed `Schema.Table`), the CLI can leave schemas out with `--exclude-schemas` and ends with a per-schema count, and a broken `project_config/` ends the server and `scripts/prompt_budget.py` in one readable line instead of a traceback, with every duplicate YAML key reported at once. The admin panel no longer leaves a card blank after a 401 or 403, and the preflight's query-timeout probe no longer logs the timeout it expects as an error.

### Added

- **`--exclude-schemas` on `python -m database.schema_inspector_cli` (PR #171).** A warehouse with many staging, replay and scratch schemas had no way to leave them out of the drafts except by listing every other schema in `--include-schemas`. `--exclude-schemas "*_Stg,Replay,Alerts"` takes comma-separated `fnmatch` patterns, matched against schema names without regard to case, applied in the inspector's table discovery next to `include_schemas`; when both are given, exclusion wins. A run now ends by printing to stderr the table count per schema, a line saying the drafts are for a first install only (an existing deployment adds tables with `python scripts/sync_schema.py --add-tables 'Schema.*'`, which keeps the curated text), and a warning when more than 150 tables were written. `docs/deployment-runbook.md` §2.3 and §16.3 and `docs/fa/getting-started.md` §0.3.2 say the same. Tests: `tests/test_schema_inspector.py`.

### Fixed

- **The schema inspector no longer writes a `schema.yaml` draft that the loader refuses (PR #171).** The draft keyed every table by its bare name and gave none a `db_schema`, so a warehouse that repeats a table name across schemas produced a file with a duplicate mapping key for each repeat (about 70 in the deployment that hit this). Every table now carries `db_schema` when its schema is known; a name unique in the run stays bare, and a name that occurs in more than one schema is keyed `Schema.Table` for every occurrence, the convention of `docs/design/TABLE-NAMES.md`. The `relationships:` entries in the draft and the entity names in the `entities.yaml` draft are spelled the same way, and a run with no foreign keys writes `relationships: []` instead of a null the validator rejects. Tests: `tests/test_schema_inspector.py` (a snapshot with the same table in two schemas passes `validate_schema_yaml_text`; a unique name stays bare).
- **A YAML file with repeated keys now reports all of them in one error, not only the first (PR #171).** `safe_load_strict` raised on the first repeated key, so a file with seventy of them took seventy runs to fix. It now reads the whole document, at every mapping level, and raises once: a single repeat keeps the message `duplicate key 'X' (first on line A, again on line B)`; several give `N duplicate keys: 'X' (lines A, B, C); ...` in file order, at most 20 and then `... and M more`. Callers that wrap the error (`knowledge.config_loader.load_yaml`, `validate_schema_yaml_text`, the preflight's `project_config/ loads`) pass it through unchanged. Tests: `tests/test_yaml_loading.py`.
- **A broken `project_config/` now ends `uvicorn api.server:app`, `python -m api` and `scripts/prompt_budget.py` in a readable message, not a traceback (PR #171).** `security.sql_guard` binds the schema allowlist at import time, so a bad `schema.yaml` (or any file the import chain reads) crashed those programs with a forty-line traceback. `core/config_preflight.py` now runs the real loaders for those files ahead of the imports that read them: `prompt_budget.py` prints `cannot load project_config: <message>` and exits with code 2, as `sync_schema.py` does; the server prints that line, one pointing to `python scripts/verify_deployment.py` and `nothing was started`, and exits with code 2. It stays fail-fast (there is no degraded mode), and nothing changes when the configuration is valid. Tests: `tests/test_config_preflight.py`.
- **An admin card is never left blank after a 401 or 403 (PR #171).** `refreshOne` in `web/admin/main.js` cleared the card, then for a 403 that is not a capability gap and for a 401 only raised the page-level banner or notice, so the card stayed an empty box. Both now also write a sentence into the card, in the `admin-loading` style: the key has no admin access for this section (see the banner), or the server did not accept the key. Tests: `tests/web_ui/test_web_ui_admin_card_errors.py`.
- **The preflight's query-timeout probe no longer logs the timeout it expects as an error (PR #171).** On a `nolock: true` source `check_query_timeout` first tried to add `WITH (NOLOCK)` to its `WAITFOR DELAY` statement (logging `WITH (NOLOCK) was not added (reason: parse_error)`) and then logged the expected timeout as `SQL execution failed on data source ...` with the full statement, once per data source, while the check passed. `execute_sql` gains a keyword-only `probe` flag that only this check sets: the statement is sent as written and its failure is logged at DEBUG. Real queries keep their hints and their error line. Tests: `tests/test_verify_deployment.py` (`TestCheckQueryTimeoutIsQuiet`).

### Upgrading

- **Nothing to change on a working deployment (PR #171).** A `project_config/` that loads today loads the same way, and no setting, file format or `requirements.lock` pin changed. Run `python scripts/verify_deployment.py` before restarting, as for any upgrade.
- **If `project_config/schema.yaml` (or `entities.yaml` / `aliases.yaml`) was replaced with a `schema_inspector` draft, restore your curated files (PR #171).** Copy them back from the backup taken before the replacement, and add the tables you want from the databases with `python scripts/sync_schema.py --add-tables 'Schema.*'` instead; it keeps the text you wrote. The inspector's drafts are for a first install only: every description is a placeholder, column notes may quote values read from live rows, and the `entities.yaml` and `aliases.yaml` drafts carry no curated aliases (their `ring_aliases` are empty), so replacing working files with them discards the curated text the prompt and the retrieval depend on. For a new install the drafts now load without duplicate-key errors (see Fixed), but they still need the review described in the runbook (§2.3) before they go into `project_config/`.

## [6.9.2] — 2026-10-10

The preflight check `Schema structure matches the databases` and `sync_schema.py --check` no longer fail on a `schema.yaml` written before 6.9.0 just because its column types are not recorded yet, and a failure is never a row of zeros: it names what differs, or says plainly that the counts do not cover it.

### Fixed

- **A `schema.yaml` written before 6.9.0 no longer fails the preflight check `Schema structure matches the databases` (and `sync_schema.py --check`) when nothing in the databases changed (PR #169).** Such a file has no `column_types:` map, so every column the database has landed in the sync plan's `recorded` list, `changed` was true, and the check failed with a message whose counts were all zero (`0 column(s) missing from schema.yaml; 0 column(s) and 0 table(s) no longer in the database; 0 column type(s) differ; 0 datasource: line(s) to set`), because that message never counted `recorded`. Every installation upgrading from 6.8 or earlier hit it on its first preflight. A column type that is merely not recorded yet is no longer drift: when recording types is the only edit a sync would make, the check passes and says `N column type(s) not recorded yet -- run python scripts/sync_schema.py once to record them (review schema.synced.yaml, replace schema.yaml), so later type changes are detected`, and `--check` exits 0 with the same note. `SyncResult.only_records_types` decides this from the plan by rendering the synced text with the recorded types left out and comparing it with the input, so a column to add, a differing type, a marker, a `datasource:` line, `--prune` and `--add-tables` all still count as drift. A failure adds `N column type(s) to record` to its counts, lists `--prune` and `--add-tables` effects when they are what differs, and a changed file that none of the counts explains now says `schema.yaml differs from the synced text in a way the drift counts do not cover -- run python scripts/sync_schema.py --dry-run to see it` instead of a row of zeros; `--check` prints the same counts. A plain `sync_schema.py` run still records the types and writes `schema.synced.yaml` as before. `docs/deployment-runbook.md` §3.1, §16.3, §17 and the Persian guide are corrected. Tests: `tests/test_schema_sync.py`, `tests/test_verify_deployment.py`, `tests/test_sync_schema_script.py`.

### Upgrading

- **`Schema structure matches the databases` now passes on a `schema.yaml` that records no column types (PR #169).** An installation coming from 6.8 or earlier sees the check pass with a `N column type(s) not recorded yet` note instead of failing; run `python scripts/sync_schema.py` once (runbook §17 step 14) to record them, so later type changes are detected. Nothing else needs doing.

## [6.9.1] — 2026-10-09

Every command-line program now writes UTF-8, so Persian output no longer crashes a Windows console or a redirected or piped stream: one shared helper, `core.console.use_utf8_console`, is called from every entry point's `__main__` block, and `scripts/release_notes.py` keeps its own standalone copy for the release workflow.

### Changed

- **docs: the upgrade path from 6.7 to 6.9.1 is complete.** `docs/deployment-runbook.md` §17 and the Persian guide's part 4 are now "Upgrading from 6.0 to 6.9.1", with ordered steps for the 6.8.0, 6.9.0 and 6.9.1 **Upgrading** notes (including that the first preflight on 6.9.x fails `Schema structure matches the databases` until `schema.yaml` is synced once, and that `sync_schema.py` arrived in 6.9.0, not 6.8.0). New runbook §18.5 and Persian §5.4 list the eleven `RETRIEVAL_*` settings and `eval.cli recall`; runbook §2.1, §2.2, §3.1 and the Persian guide are corrected for 6.8.0 (the preflight reads all ten `project_config/` files; the setup wizard follows `OPENAI_*`, obeys `LLM_ALLOW_REMOTE`, masks the connection string and resumes). `.env.example` and the README configuration table gain the `RETRIEVAL_*` settings; the README quick start gains `sync_schema.py`, its project structure the 6.8.0 to 6.9.0 modules, and its tests section `requirements-dev.txt` and `ruff check .`. No code changed.

### Fixed

- **Every command-line program now prints Persian on a Windows console (PR #167).** Reports, prompts and error messages carry Persian questions, table names and aliases; when output is piped or redirected, a console on a code page such as cp1252, cp1256 or cp437 cannot encode them, and `print` raised `UnicodeEncodeError` part-way through the output. Only `eval.cli` and `scripts/release_notes.py` guarded against it, each with its own copy. `core/console.py` now has the one helper, `use_utf8_console()`, which switches `sys.stdout` and `sys.stderr` to UTF-8 and keeps each stream's `errors` handler (`reconfigure` would otherwise reset `stderr` from `backslashreplace` to `strict`). Every entry point calls it first in its `if __name__ == "__main__":` block: `app.py`, `setup_project.py`, `webapp/app.py`, `database/schema_inspector_cli.py`, `eval/cli.py`, `eval/benchmarks/retrieval_synth.py`, `tests/test_stress.py` and every program in `scripts/` except the demo server. The web server (`python -m api`) is unchanged. `scripts/release_notes.py` keeps its own stdlib-only copy of the helper, because the release workflow fetches that file alone; a test keeps it standalone and in step with the shared one. `tests/test_cli_utf8_console.py` runs each program on a cp1252 console and fails when a new `__main__` block does not call the helper.

### Upgrading

- **Drop any console-encoding workaround (PR #167).** If a script or scheduled task sets `PYTHONIOENCODING=utf-8`, `PYTHONUTF8=1` or runs `chcp 65001` only so Persian output does not crash, it can go; leaving it in place is harmless. Nothing else needs doing.

## [6.9.0] — 2026-10-09

Table selection is now measured and improved (an offline recall harness, a 400-table synthetic benchmark, join-path expansion, inferred foreign keys and pruning: mean recall 0.71-0.75 to 0.95-0.97 on the benchmark); `schema.yaml`'s structure can be synced from the databases with the curated text kept, and the sync reports the prompt-size cost of what it adds; the prefix cache is warmed at start-up and the `llm` stage can be split into waiting and generating, with a `content: null` truncation fix; and ruff runs in CI, with optional absolute accuracy floors, a lazy `knowledge` import, one shared connection-string redactor and a password-masking fix.

### Added

- **`python scripts/sync_schema.py`: a reviewed, curation-preserving sync of `schema.yaml`'s structure (PR #162).** It reads every data source through the application's own engines (three `INFORMATION_SCHEMA` queries per source: tables, views, columns, data types, nullability, primary and foreign keys; names only, never a row) and writes `schema.synced.yaml` next to `schema.yaml`, which it never overwrites. The file is edited as text, so every comment, quote, description, flag and ordering survives. Rules: `datasource:` is set or repaired (a name, or a list for a table in several sources, as `assign_datasources.py` does); a column the database has and the file lacks is appended as `Name: "<type> column"  # TO BE FILLED (added by sync_schema.py)`; a column the file lists that no source has gets `# not in database (sync_schema.py)` above it and stays (`--prune` removes it, except one still named in `resolvable_columns` or `prefetchable_columns`); a recorded type that differs is corrected and reported; a table found in no source gets `# not found in any data source` (`--prune` removes it unless `relationships:` names it); a table the databases have and the file lacks is only reported, and `--add-tables PATTERN` (glob on `schema.table`, repeatable) adds the matching ones; a table with no `columns:` key is left alone. `--dry-run` prints the report; `--check` writes nothing and exits 1 when a sync would change the file (marked columns are listed in every report but do not fail it). The output is validated with `validate_schema_yaml_text` and compared with the original plus exactly the planned changes; anything else that differs is refused. Running it again on its own output changes nothing. A source whose catalogue cannot be read stops the run (exit 2).
- **`sync_schema.py` reports the prompt cost of what it adds (PR #165).** When anything would change, the report ends with a `== prompt size ==` section: per data source, the estimated static-prefix tokens of the current `schema.yaml` and of the synced text (both built by `build_static_prefix`, sized as `scripts/prompt_budget.py` sizes them), the delta, the path each takes, and how many added columns and tables still carry the `TO BE FILLED` draft description. A source the sync would move from the static prefix to retrieval (past `PROMPT_RETRIEVAL_TOKEN_BUDGET`) gets a `WARNING:` line naming it, both sizes and the budget, and pointing at `prompt_budget.py`. Advice only: exit codes and `--check` are unchanged, and without a `system_prompt.md` (or if the prefix cannot be built) the section is one line saying it was not computed. New: `prompt_engine.schema_preview.schema_text_in_effect`, `schema_data.registry.schema_in_effect` and `schema_data.sync.count_drafts`.
- **`relationships.proposed.yaml` (PR #162).** The same run proposes relationships for review, in the shape of `relationships.yaml`: every declared foreign key between two tables of `schema.yaml`, and `X_ID -> X.ID` links inferred from column names (`Order.CustomerID`, `Order.OrderDate_ID` to `Date.ID`). The source table's schema is preferred, both tables must share a data source, and a guess is dropped when the target has a composite or no key, the column types are of different kinds, or several tables match. Each entry carries `basis:` and `confidence:`. Nothing reads the file and `relationships.yaml` is never edited. `--no-relationships` skips it.
- **`column_types:`, an optional per-table map of `{column: SQL type}` (PR #162)** in `schema.yaml`, written only by the sync and validated against `columns:`. Nothing reads it at run time: the prompt and the SQL guard's allowlist are unchanged.
- **Preflight check `Schema structure matches the databases` (PR #162)** in `scripts/verify_deployment.py`: `sync_schema.py --check` as a check (FAIL with counts and the command to run; SKIP when a database cannot be read). Like `Tables are in their data source` it reads whole catalogues, so the admin panel runs it only on "deep checks". The failure text of `Tables are in their data source` now names `sync_schema.py`.
- **`database.catalogue.read_catalogue()` (PR #162)** returns tables and views with columns, data types, nullability and declared keys (three queries on SQL Server, SQLAlchemy reflection elsewhere).
- **Optional absolute accuracy floors in the evaluation release gate (PR #161).** The baseline comparison only blocks a relative drop (`EVAL_MAX_ACCURACY_DROP_PCT`), so a baseline that was already poor, or a slide in steps each under the allowed drop, never failed it. `EVAL_MIN_ACCURACY` / `--min-accuracy PCT` fails a run whose overall execution accuracy is below `PCT`; `EVAL_MIN_SOURCE_ACCURACY` / `--min-source-accuracy PCT` fails it when any data source is below `PCT`, or when the run has no per-source figures to check. Both are percentages in 0-100 (anything else is refused), are **off unless set**, need no `--baseline`, and a flag overrides the setting. A run exactly at the floor passes; a failure exits `1` and names the figure and the floor. `eval.baseline.check_absolute_floors()` evaluates one report and `ComparisonResult.floor_messages` carries the same result from `compare_to_baseline`. Documented in `docs/deployment-runbook.md` §18.2, `docs/fa/getting-started.md` §5.2 and `.env.example`. Set floors for `--live` runs only: an offline run is 100% by construction.
- **`core/redaction.py` (PR #161)**: `redact_db_url`, `url_secrets` and `scrub_secrets`, the connection-string redaction the setup wizard already had, now shared.
- **A `lint` job in CI (PR #161)** running `ruff check .` with ruff pinned to an exact version in `requirements-dev.txt` and the rule set (`E4`, `E7`, `E9`, `F`, what ruff applied implicitly) written out in `pyproject.toml`. Nothing is ignored; `tests/test_lint_config.py` fails if the pin, the rules or the job are loosened.
- **The prefix cache is warmed at start-up (`LLM_PREFIX_WARMUP_ON_STARTUP`, default `true`) (PR #163).** The static prefix per data source is byte-identical across requests so a model server with prefix caching only computes the tail, but nothing primed that cache: after a restart the first question paid the whole prefill (about a minute on a large local model). `llm/warmup.py` now sends the model one request per data source that uses the static prefix, from a daemon thread started by `lifespan`: the messages are exactly the prefix real requests start with (taken from `build_prompt_segments` itself, so it cannot drift), `max_tokens=1`, temperature 0, the same sampling fields and `LLM_EXTRA_BODY`. The server accepts requests at once, and a question that arrives mid-warm-up is served as before. The whole pass is bounded by `LLM_PREFIX_WARMUP_TIMEOUT_SECONDS` (default `180`, must be above zero), a request is never retried, and every failure is one log line holding no prompt text, response text or key. One INFO line per source gives the duration and `prompt_tokens` (and `cached_tokens` where the server reports it). It sends nothing for a source on the retrieval path, with `LLM_PROVIDER=mock`, or to an endpoint that is not trusted while `LLM_ALLOW_REMOTE` is off, which is why it can default to on. It helps only if the model server caches prefixes (`vllm serve --enable-prefix-caching`).
- **`POST /admin/llm/warmup` (`operations`) (PR #163).** The same pass on demand, whatever the start-up setting says, for after the model server restarts or a configuration change rewrites the prefix. Synchronous; one entry per source with `status`, `duration_ms` and `prompt_tokens`; `409` while another warm-up is running; recorded in the admin-action log as `llm.warmup`.
- **`LLM_STREAM_TIMINGS` (default `false`) splits the `llm` stage (PR #163).** With it on, `OpenAIBackend` sends `stream=true` with `stream_options={"include_usage": true}` and reassembles the chunks into the response the non-streaming call would have returned (content, reasoning text under whichever field the server uses, `finish_reason`, `usage`, `system_fingerprint`; proven on recorded chunk sequences and end to end against a stand-in server over both chunked and close-delimited transfer). The audit `llm` block gains `ttft_ms` (request to first generated token: the server's queue plus prefill) and `generation_ms` (first token to end), with `ttft_ms + generation_ms == total_ms`. The request timeout then also bounds the whole call. The constrained-decoding request is not streamed.
- **`reasoning_tokens` and `reasoning_tokens_estimated` in the audit `llm` block, with or without streaming (PR #163).** The server's own `usage.completion_tokens_details.reasoning_tokens` when it reports one; otherwise `completion_tokens` split by the share of characters that are reasoning, flagged as an estimate. `null` when the response shows no reasoning (not `0`). Only the count is recorded, never the reasoning text. The four new fields are in `docs/api-contract-v2.md` §6; a record written without them stays readable.
- **`scripts/analyze_audit_log.py` reports an "LLM stage split" (PR #163).** p50/p95 of `ttft_ms`, `generation_ms` and reasoning tokens, the median share of a call that was `ttft`, how many reasoning counts were estimates, and the share of completion tokens that were reasoning, each over the calls that have it. Also under `llm_latency_split` in `--json`.
- **docs: `docs/deployment-runbook.md` §20 and `docs/fa/getting-started.md` §7 (PR #163)** cover the warm-up (what it sends, how to read its log lines, vLLM `--enable-prefix-caching` and `--enable-prompt-tokens-details`, what it cannot do), the manual call, and how to read `ttft_ms`, `generation_ms` and `reasoning_tokens`. `.env.example`, the README and `technical_readme.md` list the three settings.
- **`python -m eval.cli recall --golden <file>` measures table selection offline (no LLM, no database) (PR #164).** For each golden case it resolves the tables `expected_sql` reads with the SQL guard's own resolution, runs `ContextRetriever`, and reports mean recall, the share of cases with every table retrieved, tables retrieved (mean, median, maximum), mean precision, an in-budget recall (a retrieval too large for `PROMPT_RETRIEVAL_TOKEN_BUDGET` counts as 0), per-tag figures and source-selection accuracy. Text or `--json`, deterministic; `--min-recall` makes it a gate. Runbook §18.4.
- **A synthetic retrieval benchmark: `python -m eval.benchmarks.retrieval_synth run --out DIR` (PR #164).** From a seed it generates a 400-table, two-source schema (star schemas, classifications, bridge tables, shared dimensions, same-name tables), its `project_config` and 480 English and Persian questions with reference SQL, and measures them. Generic words only. `docs/design/RETRIEVAL.md`.
- **Table retrieval recall work, measured on that benchmark (400 tables, 480 questions; recall / full-recall % / mean tables / precision / source accuracy %, main -> now, seeds 7 / 11 / 23): (PR #164)** 0.735 -> 0.966, 0.712 -> 0.951, 0.748 -> 0.971 mean recall; 42.1 -> 92.1, 37.3 -> 89.2, 41.7 -> 92.7 full recall; 9.2 -> 4.0, 8.2 -> 4.0, 11.0 -> 4.2 tables; 0.608 -> 0.678, 0.609 -> 0.680, 0.572 -> 0.656 precision; 90.8 -> 95.2, 90.0 -> 95.6, 86.0 -> 94.8 source accuracy. Bridge-table questions 0.558 -> 0.967 (English, seed 7). The parts:
  - **Column-name evidence.** A question word that is a word of a table's column name (`net weight` for `NetWeight`) now counts towards that table, weighted by rarity; key columns are skipped. Measure-only questions (fact table not named) 0.525 -> 0.925 (English, seed 7).
  - **Join-path expansion** (`retrieval/join_paths.py`). The tables that join the retrieved ones but were not named (a bridge table, the parent of a classification) are added from the foreign-key graph: `schema.yaml` relationships, `relationships.yaml`, and columns named `<Table>_ID` when neither declares the key (never written to configuration, no new join hint in the prompt). Shortest path of at most 2 keys, never through a hub, never across data sources, within the token budget. They are `RetrievalContext.join_tables`, part of `selected_tables`.
  - **Pruning** (`retrieval/pruning.py`). After the recall-favouring candidate generation, a table only its description matched, or a column match far below the best, is dropped unless it is joined to a better-evidenced table.
  - New settings: `RETRIEVAL_EXTRA_TABLES`, `RETRIEVAL_EXTRA_SCORE_RATIO`, `RETRIEVAL_JOIN_EXPANSION`, `RETRIEVAL_JOIN_MAX_HOPS`, `RETRIEVAL_JOIN_MAX_ADDED_TABLES`, `RETRIEVAL_JOIN_MAX_HUB_DEGREE`, `RETRIEVAL_INFER_RELATIONSHIPS`, `RETRIEVAL_PRUNE`, `RETRIEVAL_PRUNE_SCORE_RATIO`, `RETRIEVAL_PRUNE_CONNECT_HOPS`, `RETRIEVAL_PRUNE_CORROBORATE` (defaults in `docs/design/RETRIEVAL.md` §6).

### Changed

- **docs: `sync_schema.py` is the recommended way to produce `schema.yaml`'s structure (PR #162)** in the first-install checklist (steps 6 to 8), runbook §16.3 and §3.1, the Persian guide, `project_config.example/README.md` and the READMEs. `scripts/assign_datasources.py` stays, documented as the narrower tool (it writes only the `datasource:` lines and never adds, types or marks a column), because pipelines already call it; it now shares its matching and `datasource:` editing with the new script (`schema_data/placement.py`, `schema_data/yaml_text.py`), with unchanged behaviour. README: the preflight has 14 checks.
- **`import knowledge` no longer reads `project_config/` (PR #161).** `knowledge/__init__.py` imported its five names at the top, which read five YAML files the moment anything under `knowledge` was imported (`knowledge.config_loader` included), against every submodule's "import never fails" docstring. The names are now resolved on first access (PEP 562); the public API is unchanged (`__all__`, and `knowledge.X is knowledge.<module>.X`). Server start-up still stops on a missing or invalid file before the first request, because the modules `api.server` imports bind those names at import time; a test pins that. The setup wizard no longer needs to point the loaders at `project_config.example/` before importing them.
- **`build_prompt_segments` builds the static prefix and the per-request tail separately (PR #163)** (`PromptBuilder.build_static_suffix`) instead of concatenating them and cutting the result apart again at the prefix's length. Byte-identical output (`tests/test_prompt_segments.py`); the call now costs about 7 µs whatever the prefix size, where it cost 8.6 µs at a 5 KB prefix, 12.3 µs at 44 KB and 23.5 µs at 166 KB. The prefix itself was already built once per source and kept, and is invalidated when an applied configuration changes it; that is unchanged.
- **An alias or fact-pattern hit no longer ends the search for tables (PR #164).** `EntityRetriever` and `FactRetriever` returned only the aliased tables when any alias matched, so a question naming one aliased and one unaliased dimension lost the second, and an `always_include` table (the date dimension) vanished whenever another dimension matched. Up to `RETRIEVAL_EXTRA_TABLES` ranked tables are now added. Alias matching folds Persian/Arabic spellings and ZWNJ, and results keep configuration order instead of set order (they varied between runs).
- **The TF-IDF retriever splits on word characters (PR #164).** `customers?`, `سال؟` and `sales.Order` were single tokens and never matched a description word. `schema_data.retriever` gains `rank_tables`, `forced_tables` and `column_evidence`; `retrieve_tables` keeps its meaning and breaks score ties by table name.
- **Static-prefix prompts are unchanged (PR #164)**; only the tables the retrieval path lists differ. `RETRIEVAL_PRUNE=false`, `RETRIEVAL_JOIN_EXPANSION=false` and `RETRIEVAL_EXTRA_TABLES=0` restore the previous selection.

### Fixed

- **`python -m eval.cli` no longer crashes on a Windows console (PR #164).** Reports carry Persian questions and table names, and a console on a code page such as cp1252 cannot encode them, so `eval.cli recall` (and any subcommand printing Persian) failed with `UnicodeEncodeError` part-way through its output. The CLI now writes UTF-8 to stdout and stderr when run as a program, and the benchmark reads its child process's output as UTF-8.
- **Ruff is clean (89 findings fixed, none ignored) (PR #161).** Unused imports and variables removed, `l` renamed, two one-line loops split. Four `F821` undefined names were all one: `"PromptSegments"` in the annotations of `LLMBackend.generate_with_meta_segments` / `generate_structured` and the two provider overrides. Not a runtime bug (the annotations are strings under `from __future__ import annotations` and nothing evaluated them), but `typing.get_type_hints`, Sphinx or pydantic would have raised `NameError`; the name is now imported under `TYPE_CHECKING` (a runtime import would be circular), and a test resolves the hints.
- **A response with `content: null` is an empty answer, not the text `None` (PR #163).** A reasoning model cut off while still thinking returns `finish_reason: "length"` and `"content": null` (its thinking is in a separate field), and `OpenAIBackend` rendered that with `str(None)`. The four characters looked like an answer, so `is_truncated_empty_completion` never fired and the `LLM_OUTPUT_TRUNCATED` message, which names `LLM_NUM_PREDICT` and `LLM_EXTRA_BODY`, was replaced by a parse failure about SQL that was never written. A `null` or missing `content` is now `""` on the non-streaming and the streaming path alike, in the free-text and the constrained-decoding request (which used to raise a `TypeError` on `null`), and the reasoning-token estimate no longer works around the string. `mode="sql"` and the v2 session engine now report `LLM_OUTPUT_TRUNCATED` for such a response, as they always did for an empty string. `mode="full"` on `/query` goes through the correction loop, which has no truncation shortcut: a cut-off response there is still retried and reported as `INVALID_SQL_RESPONSE`, for `""` and `null` alike.

### Security

- **One robust connection-string redactor (PR #161).** `database/schema_inspector.py` masked passwords with a regular expression that left part of a password containing `@` or `:` visible and never touched `PWD=` inside an `odbc_connect` value, so the redacted URL in a schema snapshot could still carry a credential. `appdb/engine.py`'s `redact_url` and `DataSource.redacted_url` (used in refusal messages and `scripts/verify_deployment.py` output) had the same `odbc_connect` gap. All now use `core.redaction.redact_db_url`; the wizard keeps its `_redact_db_url` / `_scrub_secrets` names.

### Upgrading

- **Table retrieval selects differently by default (PR #164).** Eleven new `RETRIEVAL_*` settings control it (defaults in `docs/design/RETRIEVAL.md` §6); among them `RETRIEVAL_EXTRA_TABLES` defaults to `3`, and `RETRIEVAL_JOIN_EXPANSION` and `RETRIEVAL_PRUNE` default to `true`. Static-prefix prompts are unchanged; only the tables the retrieval path lists differ. `RETRIEVAL_PRUNE=false`, `RETRIEVAL_JOIN_EXPANSION=false` and `RETRIEVAL_EXTRA_TABLES=0` restore the previous selection.
- **Start-up now sends one small request per data source to the model server (PR #163).** `LLM_PREFIX_WARMUP_ON_STARTUP` defaults to `true` (the warm-up above). If the model server is shared and metered per token, or start-up must send it nothing, set it to `false`. Nothing else needs doing; `LLM_STREAM_TIMINGS` stays off until you turn it on.
- **Regenerate any schema snapshot written by `database/schema_inspector.py` (PR #161)** if its `source_url` may have been shared and quoted a password: earlier versions could leave part of the password in it. Rotate that password if it was.
- **`scripts/sync_schema.py` is now the recommended way to maintain `schema.yaml`'s structure (PR #162).** It writes `schema.synced.yaml` for review and never overwrites `schema.yaml`; `scripts/assign_datasources.py` stays for pipelines that already call it.
- **Contributors: install `requirements-dev.txt` (PR #161).** It pins `ruff` to an exact version, and CI now fails on `ruff check .`.

## [6.8.0] — 2026-10-09

A column listed in `denied_columns` can now be restricted to one table or one data source and used only as a join key; the setup documentation is rewritten around one ordered first install; the setup wizard no longer stores the database password in `.setup_log.json` and now obeys `LLM_ALLOW_REMOTE`, with its resume, regenerate and validation steps fixed.

### Added

- **Scoped, join-only column restrictions in `denied_columns` (PR #158).** Alongside the plain name (`NationalID`, unchanged: denied everywhere, every reference), an entry may be `schema.Table.Col` (join-only on that table, in any source that holds it), `Source:schema.Table.Col` (only when the query executes on `Source`) or `Source:Col` (every table of `Source` that has the column). A join-only column may appear only as one side of `a.col = b.col` inside a `JOIN ... ON`; the select list, `WHERE`, `GROUP BY`, `ORDER BY`, `HAVING`, a function argument, `ON` against a literal, `USING (...)`, a comma join's `WHERE` predicate and a `*` that expands to it are refused with the new, correctable guard reason `join_only_column` (its `subject` is the whole entry, so "Request access" removes exactly that entry). The data source is the one `database.routing.choose_datasource` picks, the same call the executor makes. Entries are parsed by the new `security/column_policy.py` and checked against `schema.yaml` and `datasources.yaml` when keys are loaded and when the admin panel saves an ACL, so a typo stops start-up (or answers 422) naming the entry. The model is told which columns are join-only by one line in the per-question part of the prompt; the cached static prefix is unchanged. The admin key editor shows and accepts the new forms.

### Changed

- **`scope_key` hashes the sorted `denied_columns` as a JSON array instead of joining them on `":"` (PR #158).** Scoped entries contain `:`, so the old join could give `("a:b",)` and `("a", "b")` the same key. Cached query results are keyed afresh once after the upgrade.
- **The value resolver and the dimension vocabulary skip join-only columns (PR #158)**, so their distinct values cannot reach the prompt.
- **docs: the setup documentation is rewritten around one ordered first install, and brought up to 6.7.0 (PR #157).** `docs/deployment-runbook.md` and `docs/fa/getting-started.md` open with "First install, in order", a 17-step list from an empty machine to a verified, running server, each step with its exact command and how to tell it worked (install, `.env`, `project_config/`, `datasources.yaml`, `scripts/assign_datasources.py` with a backup and `--check`, the columns it reports missing, per-source `nolock:`/`keywords:`/`description:`, the first API keys, `scripts/verify_deployment.py`, `scripts/prompt_budget.py`, starting the server, the golden set); the first admin key comes before the preflight because its `API key authentication` check fails while no key exists. New runbook §2.1 to §2.3 (and Persian ۰.۳.۱, ۰.۳.۲) document creating `project_config/`, the setup wizard `setup_project.py` (its options, its seven steps and the `.setup_log.json` keys, what it writes and what it does not: no `schema.yaml`, no `datasource:` per table; that it reads its own `WIZARD_*` settings and not `OPENAI_BASE_URL`, `DB_PASSWORD` or `datasources.yaml`) and `python -m database.schema_inspector_cli`, the tool that drafts `schema.yaml`, which no document mentioned. New runbook §3.1 (Persian ۰.۴) is a reference table of all 13 `verify_deployment.py` checks with what each does, what `[FAIL]` and `[SKIP]` mean (there is no warning status) and the fix, including `Tables are in their data source`. `nolock: true` is stated to be per data source, with a two-source example. `DB_APPLICATION_NAME` and a DMV query for the application's sessions and open transactions are in the Persian guide as well as the runbook. "Upgrading from 6.0 to 6.6" (§17, Persian بخش ۴) is now "6.0 to 6.7" and gains the 6.7.0 golden-set and release-gate step. New runbook §19 and Persian بخش ۶ say which files never to send when asking for help (`.env`, `api_keys.json`, raw keys, `.setup_log.json`, logs, `eval_data/`) and which are safe. README: quick start reordered so the preflight runs before the server starts, the wizard and the inspector, `harvest_golden.py`, `golden_sheet.py`, `create_db.py` and the `eval/` modules added to the project structure, the evaluation feature row updated, and the test count and coverage re-measured at 6.7.0 (5,470 tests, 94%). The tutorials and `project_config.example/README.md` point to the new sections. No code changed.

### Fixed

- **`--resume` continues from the first step `.setup_log.json` does not record as completed (PR #159).** It used to skip only files that already existed and ran every step again, the model calls included. The log now keeps the aliases, rules and examples each LLM step produced, with the language, the provider and a fingerprint of the schema. A resumed run reconnects (the connection string is never stored) and re-reads the schema, reuses the stored results of completed LLM steps while the schema and language are unchanged (never ones made by the mock LLM), and writes only the files that do not exist; when steps 1 to 6 are all recorded and every file exists it runs only the validation, without the database or the model. A run without `--resume` starts a new log, and `--dry-run` no longer writes one (it also no longer marks step 6 done with nothing written).
- **The wizard's "Regenerate" choice regenerates (PR #159).** No caller passed a regeneration function, so it showed the same text again. Step 6 now re-runs step 4 for `business_rules.yaml` and step 5 for `examples.yaml`; for `entities.yaml` and `aliases.yaml`, which are both built from the step 3 aliases, it re-runs step 3 for every table without the per-table questions. A failure is reported and the current version kept. The choice is not offered for `relationships.yaml` (built from the schema) or when there is no LLM.
- **Step 7 of the setup wizard reports missing `project_config/` files instead of crashing (PR #159).** Importing anything under `knowledge/` reads five files of the default `project_config/` on the spot, so on a fresh checkout step 7 (and the interactive validation) died with a `ConfigNotFoundError` traceback. The wizard now validates the files it wrote and lists the required files still missing, with the command that copies each template from `project_config.example/`; it does not copy them (the application has no fall-back to the sample configuration either) and does not say "Setup complete" while files are missing. The ten required names are one list, `core/project_config_files.py`, pinned by a test to the loaders.
- **The setup wizard's LLM follows the application's settings (PR #159).** An empty `WIZARD_LLM_BASE_URL` (as in `.env.example`) silently meant `https://api.openai.com/v1`, and the shipped `WIZARD_LLM_MODEL=gpt-oss-20b` shadowed `OPENAI_MODEL`. Endpoint, model and key now come from the flag, then `WIZARD_LLM_*`, then `OPENAI_BASE_URL` / `OPENAI_MODEL` / `OPENAI_API_KEY`; the wizard prints one line naming the provider, model, endpoint and where each came from (never the key), a local endpoint no longer needs a key, and a fall-back to the mock LLM because the endpoint did not answer is announced loudly. `.env.example` leaves `WIZARD_LLM_MODEL` and `WIZARD_LLM_BASE_URL` empty and explains this.
- **`scripts/verify_deployment.py` checks all ten required `project_config/` files (PR #159).** "project_config/ loads" loaded six, so a missing or invalid `retrieval_hints.yaml`, `session_policy.yaml`, `memory_policy.yaml` or `system_prompt.md` passed. It now loads all ten through the application's own loaders, names every missing file at once, resolves a relative `PROJECT_CONFIG_DIR` against the repository root as the application does, and fails on an empty system prompt.

### Security

- Join-only hides a column's value from the output, not its existence: a join on a filtered foreign key can still probe it, and hiding a primary key whose foreign key is visible elsewhere is cosmetic. Restrict the foreign keys as well. An unqualified reference counts as restricted when any table in its `SELECT` is restricted for that name. An entry name cannot contain `.` or `:`; a legacy entry that did is now read as a scoped one and fails start-up if it names nothing real. See `docs/deployment-runbook.md` and `docs/design/DESIGN.md` D14 (PR #158).
- **The setup wizard no longer stores the database password (PR #159).** `.setup_log.json` recorded the connection string under `db_url_redacted` through a `'re' in dir()` test that is always false inside a function, so it held the URL exactly as typed. It is now redacted through SQLAlchemy's URL parser, which also masks a password in the query string or inside `odbc_connect` and a password containing `@`, `:` or `/`, with a masked-by-pattern fallback for a string the parser rejects. The same helper masks the `Source:` line in `entities.yaml`, and driver errors shown on a failed connection or schema inspection are scrubbed of the URL and the password. A `.setup_log.json` written by an earlier version may hold the password: delete it, and rotate the password if the file was shared or committed.
- **The setup wizard obeys `LLM_ALLOW_REMOTE` (PR #159).** It sends table and column names and sample column values to the model, and never consulted the gate the application applies. It now uses the application's definition of a remote endpoint (`LLM_TRUSTED`, else loopback, private-network and `*.local` hosts are local) and refuses a remote one, before it touches the database or the endpoint, unless `LLM_ALLOW_REMOTE=true`.

### Upgrading

- **Delete any `.setup_log.json` written by an earlier version, and rotate the database password if that file was shared or committed (PR #159).** Earlier versions stored the connection string in it exactly as typed, password included. The wizard now writes a redacted URL.
- **The query cache is invalidated once (PR #158).** `scope_key` now hashes `denied_columns` as a JSON array, so every cached result is keyed afresh after the upgrade; nothing needs to be cleared by hand.
- **Scoped `denied_columns` entries are validated at startup (PR #158).** A mistyped `schema.Table.Col`, `Source:schema.Table.Col` or `Source:Col` entry now fails startup loudly instead of silently restricting nothing; fix the entry it names. Plain names are unchanged.
- **`WIZARD_LLM_MODEL` and `WIZARD_LLM_BASE_URL` in `.env.example` now default to empty and fall back to `OPENAI_MODEL` and `OPENAI_BASE_URL` (PR #159).** If you copied the old `.env.example` values (`gpt-oss-20b`, and an empty base URL meaning `https://api.openai.com/v1`) into your `.env`, empty them to follow the application's own settings; the wizard prints one line naming the endpoint and model it uses.

## [6.7.0] — 2026-10-05

A real evaluation set can now be built from real usage and used to gate an upgrade; CI tests the pinned dependency set as well as the newest releases; the documentation is brought up to date.

### Added

- **A real evaluation set can be built from real usage and gated on before an upgrade (PR #155).** `python -m eval.cli run --live --reference live` executes each case's `expected_sql` in the same run and compares the two results as execution accuracy (`eval/compare.py`: rows as a multiset, column names ignored and column order as selected, row order only for a top-level `ORDER BY` with `TOP`/`OFFSET`, numbers within `--float-tolerance` (default `1e-6`) but integers exactly, `NULL` equals `NULL`), so a warehouse that changes daily no longer turns a stale recorded fingerprint into a false failure; the default `--reference stored` keeps the old fingerprint comparison, `--reference live` requires `--live`, and a baseline can only be compared with a run that used the same reference. `scripts/harvest_golden.py` writes a seeded, stratified sample (data source, outcome, language; default 150) of de-duplicated candidates from the audit log to `eval_data/candidates.jsonl` as `pending_review` cases, printing counts only unless `--include-examples`; `scripts/golden_sheet.py export`/`import` round-trips them through a UTF-8-with-BOM CSV for Excel review (`eval_data.example/golden_review_template.csv`), validating every SQL with the guard and reporting problems by row; `python -m eval.cli verify [--accept]` runs each reviewed case's SQL read-only through the application's executor, records `expected_rows`/`expected_fingerprint`, and only with `--accept` makes it `active` (atomic write, `.bak` kept). `GoldenCase` gains an optional `datasource` and the statuses `pending_review` and `reviewed` (the runner skips every status but `active`); the report adds per-source execution accuracy and source-selection accuracy, and baselines written before these fields still load. See `docs/deployment-runbook.md` §18, `eval_data.example/README.md` and `docs/fa/getting-started.md` §۵.

### Changed

- **CI now tests the pinned dependency set that deployments install, as well as the newest releases (PR #152).** The `test` job in `.github/workflows/ci.yml` gained a `deps` matrix dimension: the existing legs still install `requirements.txt` (floors only, so the newest releases, which warns early when an upstream release breaks us), and a second set of legs, named `<os> / Python <version> / locked deps`, install `requirements.lock` across the same three operating systems and Python 3.11, 3.12 and 3.13 (every pin has a wheel on all nine combinations). They run the same suite, doctests and offline evaluation run; `webapp/requirements.txt` and pytest are installed under the lock as a constraint so nothing can move a pin. Until now CI never ran the tree a deployment installs (sqlglot, whose AST `security/sql_guard.py` decides from, is pinned at 30.18.0 while CI ran 30.21.0). The existing check names are unchanged; the pip-audit, coverage and upload steps still run once, on the Linux 3.12 floors leg.
- **The documentation is brought up to 6.6 and reorganised around one place per topic (PR #153).** `docs/deployment-runbook.md` gains an install step (`pip install -r requirements.lock`, with the reason, and what CI tests: the pinned lock and the newest floors), §16 rewritten as the operator's ordered procedure for several data sources (describing the sources, `scripts/assign_datasources.py`, `keywords:`, preflight, `scripts/prompt_budget.py`, `nolock`, reading the schema-drift and vocabulary cards, reading `datasource_selection`) and a new §17, "Upgrading from 6.0 to 6.6", which sequences the **Upgrading** notes of 6.0.1 to 6.6.1 into one checklist; `docs/design/DATASOURCES.md` stays the one full explanation and gets four decision records (DS1 to DS4), and `docs/design/DESIGN.md` gets D13 for the display-only SQL layout. The Persian guide and tutorial cover the same operator path, with the duplicated long explanations in the README, the tutorials and the guides replaced by short pointers. Corrected against the code: the prompt has a static path (whole schema, the default) and a retrieval path (over `PROMPT_RETRIEVAL_TOKEN_BUDGET`) rather than always being scoped; the guard is parser-based, not a keyword list; the tutorials' domain-authoring steps are the YAML files in `project_config/`, not `knowledge/*.py`; `/query` responses, `/health`, rate-limit and cache descriptions, the admin panel's twelve sections and their capabilities, the `DIMENSION_VOCABULARY_WARM_ON_STARTUP` default (`true`), `DEFAULT_TOP_N`, the CLI banner and log file names, and the test count. `PROMPT_RETRIEVAL_TOKEN_BUDGET` is now in `.env.example` and the README's configuration table; `SCHEMA_CACHE_TTL`, which nothing reads, is removed from `.env.example`; and `AUTO_DISCOVER_SCHEMA` is documented as read only by `database/relationship_map.py`, which nothing in the server imports. No code changed.

## [6.6.1] — 2026-10-04

A template for the `API_KEYS_FILE` key array.

### Added

- **`project_config.example/api_keys.example.json` is a template for the `API_KEYS_FILE` file (PR #150).** Copy it to `project_config/api_keys.json` (git-ignored) and replace each `key_sha256` with the digest `scripts/issue_api_key.py` prints. It holds an analyst entry with a `denied_columns` example and an admin entry with `admin`, `operations` and `security`, which together show every field the parser reads. Both digests are placeholders, so a copy that was not edited is refused at start-up (`key_sha256 must be a 64-character SHA-256 hex digest`) instead of loading keys nobody holds. `.env.example`, the deployment runbook, `project_config.example/README.md` and the Persian getting-started guide point to it; `tests/test_api_keys_file.py` pins its fields and the refusal.

## [6.6.0] — 2026-10-04

`scripts/prompt_budget.py` measures what each data source's prompt really costs and prints the `PROMPT_RETRIEVAL_TOKEN_BUDGET` to set.

### Added

- **`scripts/prompt_budget.py` shows what each data source's prompt really costs and which `PROMPT_RETRIEVAL_TOKEN_BUDGET` to set (PR #148).** Run it from the repository root: it builds each source's static prefix exactly as the server does (the system prompt from `PROJECT_CONFIG_DIR`, `build_static_prefix(system_prompt, source)`; the one implicit source when there is no `datasources.yaml`) and reports its characters, table count, the heuristic estimate, and the path it takes under the current budget. Unless `--no-model` is given it also sends each prefix to the endpoint the router uses for SQL generation as one chat completion with `max_tokens=1` and reads `usage.prompt_tokens`, using the application's own request body, headers and `LLM_EXTRA_BODY` and its data-governance gate (an untrusted endpoint is sent nothing unless `LLM_ALLOW_REMOTE` is true); this warms the model server's prefix cache for each source. The context length comes from `--context-length N` or `GET {base}/models` (`max_model_len`, `context_length` or `context_window`), else it is unknown. With it the script checks per source that real tokens (or the estimate times the largest ratio measured elsewhere, or times `--assumed-ratio`, default 1.15, labelled as such) + `--question-room` (default 2000) + `LLM_NUM_PREDICT` fit the window, leaves a source that does not fit out of the recommendation with a warning (it should stay on the retrieval path) and exits 1, and recommends the budget as the largest estimate among the sources that fit, plus `--headroom` percent (default 10), rounded up to a multiple of `--round` (default 500), with the exact `.env` line and each source's path now and after. If the endpoint is unreachable or returns no usage a source's count is shown as `unavailable` and the run continues. `--json` prints one JSON document; no API key or URL credential is printed; exit 2 means the configuration could not be loaded or an option is invalid.

### Upgrading

- **With several data sources, run `python scripts/prompt_budget.py` once after upgrading (PR #148)** and set the `PROMPT_RETRIEVAL_TOKEN_BUDGET` it prints in `.env`; it needs the model endpoint reachable for measured numbers (`--no-model` works offline from the estimate).

## [6.5.0] — 2026-10-04

**Issues fixed:** #130, second part (one table that exists in two databases: `datasource:` as a list).

A data source can read its tables `WITH (NOLOCK)`, a table in `schema.yaml` may now live in several data sources, `scripts/assign_datasources.py` writes the `datasource:` lines for you, schema drift says where a table really is, and with several sources each question is routed to one of them so the model sees that source's whole schema; the SQL shown in the UI now has a fixed, aligned layout.

### Added

- **`datasource:` accepts a list, for a table that exists in several data sources (PR #146).** `datasource: [sales, inventory]` says the same table, with the same shape, is in each (a date dimension replicated into both databases is the usual case); a plain name works as before. The list must be non-empty and its names distinct, and every name is checked against `datasources.yaml` wherever names are checked: at start-up, in the admin panel's `schema.yaml` draft validation and by `verify_deployment.py`. A statement now runs on a source that has **every** table it reads: the default source if it is one of those, otherwise the first in `datasources.yaml` order. A statement that reads only a shared table runs on the default source, and so do the value resolver and the dimension-vocabulary prefetch for it (one copy is read, not each). The SQL guard applies the same rule as the executor (`database.routing.choose_datasource`), so they cannot disagree. A statement is refused as `cross_datasource` only when no source has all its tables, and the message still names every source with its tables, now also listing which tables are available in several. The prompt's schema block prints `Data source: sales, inventory` for a shared table.
- **`scripts/assign_datasources.py` works out each table's `datasource:` from the databases (PR #146).** Run it from the repository root: it reads the tables, views and columns of every data source through the application's own read-only engines (two `INFORMATION_SCHEMA` queries per source, names only), matches every `schema.yaml` table by schema and name, and writes `schema.with_datasources.yaml` next to `schema.yaml` (`--output PATH` to choose another file; `schema.yaml` itself is never overwritten) with every original line and comment kept and one `datasource:` line under each table key: a name, `[A, B]` for a table found in both, and `# not found in any data source` under a table found in neither. An existing `datasource:` entry is replaced, not duplicated. The output is validated with the project's own `validate_schema_yaml_text`, and the script refuses to write a file that fails it or differs from `schema.yaml` in anything but `datasource`. The report lists, per source, the tables found only there, the shared tables, the tables found nowhere, the tables whose current `datasource:` disagrees with what was found, and the columns `schema.yaml` lists that the database does not have. `--check` writes nothing and exits 1 if any `datasource:` disagrees; exit 2 means a catalogue could not be read or the file could not be produced.
- **Schema drift and `verify_deployment.py` say where a misplaced table is (PR #146).** When every column of a table is missing from the source `schema.yaml` assigns it to and another source has the table, `GET /admin/schema-drift` returns it in a new `misplaced_tables` list with a hint such as `stock_dim.Broker: not in sales, found in inventory — set datasource: inventory`, and the admin panel's schema-drift card lists it with the value to write. It costs one `INFORMATION_SCHEMA.TABLES` query per other source, only when such a table exists, and is cached with the rest of the drift result. `verify_deployment.py` has a new check, `Tables are in their data source`, that fails with the same hints (skipped with one data source); the admin panel runs it only under "deep checks", because it reflects every catalogue.
- **A data source can read its tables `WITH (NOLOCK)` (`nolock: true` in `datasources.yaml`) (PR #146).** For warehouses where the DBA requires the hint on every read. It is off by default and set per source, in the structured and the `url_env` form; anything but a YAML `true` or `false` is refused naming the source, and so is `nolock: true` when `SQL_DIALECT` is not `tsql` (table hints are T-SQL). Just before a statement is sent, `database.executor` passes it through the new `database.table_hints.add_nolock_hints` when its routed source has the flag, which inserts ` WITH (NOLOCK)` after each physical table reference, after the alias when there is one, in `FROM`, every `JOIN`, subqueries, CTE bodies and each branch of `UNION`/`EXCEPT`/`INTERSECT`. Nothing else in the text changes: sqlglot only locates the references and the hint is inserted at that offset, so literals, comments, `?` placeholders, `TOP`, `OFFSET`/`FETCH` and the SQL as the model wrote it stay as they were. CTE names, derived tables, table-valued functions, `#temp` tables, `@table` variables, `INFORMATION_SCHEMA`/`sys` objects and a table that already has a `WITH (...)` hint are left alone. The rewritten text is parsed again and checked; if that fails, or the statement cannot be parsed, it runs unchanged and one warning per reason is logged. The audit trail's `generated_sql` stays the validated, unhinted SQL; only the text sent to the server differs. `NOLOCK` allows dirty reads (uncommitted rows, and occasionally a row twice or missing), so turn it on only where it is required. `project_config.example/datasources.example.yaml` documents `nolock` and the `APP` ODBC option (the `program_name` a DBA sees in `sys.dm_exec_sessions`), and `DialectProfile` gains `supports_table_hints`.
- **With several data sources, each question is routed to one source and the model is shown that source's schema (PR #146).** A statement can only run on one source, so the prompt no longer describes every table of every source. The source is chosen before the prompt is built, with no model call (`retrieval/source_selector.py`): first by `keywords:` in `datasources.yaml`, then, for a follow-up turn of a conversation, by the previous turn's source, then by the tables retrieval selected and the tables whose warehouse values matched (a table in several sources counts for each, with a lower weight), then by the default source. The prompt for the chosen source holds only the tables that live in it (shared tables included), only the relationships whose two ends are both among them, and only the few-shot examples whose SQL reads only such tables (an example whose tables cannot be determined is kept); business rules and metrics stay whole, and `Data source: <name> — <description>` is printed once above the schema. `build_static_prefix`, `static_prefix_token_estimate` and `should_use_static_prefix` take the source, so each source has its own cached, byte-identical prefix and its own budget decision: a source whose prefix fits `PROMPT_RETRIEVAL_TOKEN_BUDGET` uses it, a larger one uses the retrieval path restricted to its tables, and when retrieval finds no table that path now shows every table of the chosen source, not of the deployment. Both engines build the prompt through `llm.router.build_prompt_segments(..., source=)`. A deployment with one data source is unchanged: prompts, prefixes, cache keys and the estimate are byte for byte what they were (pinned by `tests/test_source_prompts.py`).
- **`keywords:` in `datasources.yaml` (PR #146).** An optional list of words or phrases, Persian or English, that mark a question as being about a source (`description:` already existed and is now also printed above the source's tables in its prompt). A phrase matches as whole words after the folding the retrieval layer uses everywhere (Arabic `ي`/`ك` to `ی`/`ک`, Persian and Arabic digits to `0-9`, ZWNJ removed, case ignored), so `stock` does not match `stockholder`. A value that is not a list of non-empty strings, or lists a phrase twice (two spellings that fold to the same text count as twice), is refused at start-up naming the source. `database.datasources.datasource_keywords()` and `datasource_descriptions()` read them without touching credentials.
- **One retry when the model answers `OUT_OF_SCOPE` for the chosen source (PR #146).** If another source exists, the request is retried once with the next candidate (ordered by strength of evidence): one extra model call at most, never more, outside the correction budget, with the correction history dropped because it describes the other source's tables. A second `OUT_OF_SCOPE` is returned as before, and a deployment with one source never retries. `llm.source_routing.generate_with_source_fallback` is the one helper `SQLAgent` (`/query`, result, full and SQL-only modes) and `TurnEngine` (`/v2` sessions) both call.
- **The audit record gains `datasource_selection` (PR #146).** `{"chosen", "reason", "candidates", "fallback_from"}`: the source whose tables the model last saw, the signal that ordered the candidates (`keyword`, `session`, `retrieval` or `default`), every source in fallback order, and the source that declined when the request retried. `None` with one source and for a result served from the cache; additive, like `datasource`, which still records where the generated SQL routes (the guard and router stay authoritative, and no new refusal is added when the two differ). `prefix_cache_hit` in the audit `llm` block is now measured against the chosen source's own prefix estimate. With several sources one INFO line per source at start-up (or first use) says its table count, its static-prefix estimate, the budget and the path its prompts take.
- **A conversation keeps its source between turns (PR #146).** `session.store.TurnMemory.datasource` records the source a turn was answered from and is persisted with the turn (only when set, so a single-source deployment stores exactly what it did). A follow-up turn starts from it unless the question's keywords point elsewhere; a refinement composed over the previous turn's SQL stays on that source; a question the session classes as new is judged on its own.

### Changed

- **Schema drift checks a shared table in each of its sources (PR #146).** A column missing from one copy is reported as `Table.Column [source]`, so the copies are no longer merged; tables in one source keep their plain `Table.Column`. A copy of a table that `schema.yaml` describes only for another source is no longer merged into that table's columns.
- **`TableDefinition.datasource` is a tuple of names, and `get_table_datasource_names()` returns tuples (PR #146)** (`()` for a table with no `datasource:` key, where both used to be a string). `database.datasources.table_datasources()` still returns `{table: one source}`; `table_datasource_sets()` is new and returns every source of every table, and `database.routing.group_tables_by_datasource()` lists a shared table under each of its sources.
- **`PROMPT_RETRIEVAL_TOKEN_BUDGET` applies per data source when several are configured (PR #146).** The default (6000) is unchanged. It is compared with each source's own prefix estimate, not the whole schema's. The estimator is still `len(text) // 4`, which undercounts Persian text by roughly 15%, so leave that margin when sizing it; `docs/design/DATASOURCES.md` ("Choosing a source per question") and `docs/deployment-runbook.md` §16 say how.
- **`build_static_prefix(system_prompt, source=None)`, `static_prefix_token_estimate(system_prompt, source=None)`, `should_use_static_prefix(system_prompt, source=None)`, `PromptBuilder.build(..., source=None)`, `SchemaRegistry.build_schema_context(..., source=None)` and `llm.router.build_prompt_segments(..., source=None)` take an optional data source (PR #146).** `None` is what every call site did before, and with one source configured a name is ignored. `SchemaRegistry.tables_for_source()` is new. `SQLGenerationResult` gains an optional `datasource_selection`.
- **`Turn.sql_display` is laid out in the house style instead of sqlglot's default (PR #146).** The new `security/sql_format.format_sql` writes `SELECT` alone on its line (`SELECT TOP (100)`, `SELECT DISTINCT`), one item per line with a leading comma (the first indented five spaces, the rest four and a comma), the `AS` of every alias in one column two spaces after the longest aliased expression (a `CASE`, or an expression longer than 60 characters, does not stretch it), a blank line before `FROM`, and one line per join (`INNER JOIN`, `LEFT JOIN`, `RIGHT JOIN`, `FULL JOIN`, `CROSS JOIN`, `CROSS APPLY`, `OUTER APPLY`, always in full and in upper case) with the `AS`, `ON` and `=` columns aligned. `WHERE`, `HAVING` and multi-condition `ON` continue with right-aligned `AND` / `OR`, and derived tables, `IN` / `EXISTS` / scalar subqueries, CTE bodies and the branches of `UNION` / `EXCEPT` / `INTERSECT` are laid out by the same rules one level in. Expression text still comes from sqlglot's T-SQL generator (with `IS NOT NULL` and `NOT IN` written as a person writes them), and identifier quoting, `N'...'` literals, numbers and `?` placeholders are shown as parsed. The result is parsed again and compared with the parse of the input; if the trees differ, if the statement has a comment or a construct the layout does not handle (`MERGE`, `PIVOT`, comma joins, `OPTION`, `INTO`, `ROLLUP`, ...), or if anything raises, `pretty_sql` gives sqlglot's pretty output as before and, failing that, the input. Several statements in one string, which `pretty_sql` used to cut down to the first, now come back unchanged. `Turn.sql`, the text that is validated, executed and audited, is never touched, and the web UI needs no change: it shows a multi-line `sql_display` as it is, in a `white-space: pre` block.
- **Displayed SQL is never a rewrite of what ran, a refused statement is laid out too, and the SQL shown while a turn streams is the final one (PR #146).** `pretty_sql` (every dialect) now uses sqlglot's pretty printer only when its output parses back to the same tree as the input and keeps the case of every function name sqlglot does not know; otherwise the input comes back unchanged, so `DATEDIFF(day, a, b)` is no longer shown as `DATEDIFF(DAY, CAST(a AS DATETIME2), CAST(b AS DATETIME2))` nor `FROM dbo.fn(1)` as `FROM dbo.FN(1)`. `GuardVerdict.rejected_sql_display` is `pretty_sql` over `rejected_sql` (`null` when `rejected_sql` is, equal to it when nothing could be laid out) and is serialised with it, in the turn and the `sql` and `done` events and the stored session; the «دیدن SQL» reveal shows it when present and falls back to `rejected_sql` through the client formatter, which is never altered. The `sql` stream event gains `sql_display`, so the first paint is the text `done` brings instead of the client formatter's layout; a client that ignores the field is unaffected. The generative path now sets `Turn.sql_display` (from the executed SQL, as the refinement path already did), where it was `null` and the client formatted `sql` itself, so the copy button on such a turn copies the laid-out statement.

### Upgrading

- **Nothing changes for a deployment with one data source, or whose tables each name one source or none (PR #146):** prompts, routing, errors and drift output are as before.
- **With several sources and no `datasource:` lines yet (PR #146),** run `python scripts/assign_datasources.py`, review `schema.with_datasources.yaml`, and replace `schema.yaml` with it (it needs a restart like any `schema.yaml` change to the allowlist). `--check` in a deploy pipeline keeps the assignments honest afterwards.
- **With several sources (PR #146),** add `description:` and `keywords:` to each source in `datasources.yaml` (optional; routing works from the retrieved tables without them), restart, and read the start-up lines `Prompt path for data source ...`: a source over `PROMPT_RETRIEVAL_TOKEN_BUDGET` takes the retrieval path restricted to its tables, so raise the budget (leaving ~15% for Persian text) or move tables to another source if you want its cacheable prefix. Audit records now carry `datasource_selection`; a reader that does not know it can ignore it.

## [6.4.1] — 2026-10-03

python-dotenv moves to 1.2.4 so a `.env` saved with a byte-order mark loads its first variable.

### Changed

- **python-dotenv is 1.2.4 in `requirements.lock` and `>=1.2.3` in `requirements.txt` and `webapp/requirements.txt` (PR #144).** Releases from 1.2.3 on strip a leading UTF-8 byte-order mark themselves, where 1.2.2 read it as part of the first variable's name and silently lost that variable. Windows editors often save a `.env` that way, so a `.env` saved as UTF-8 with BOM now loads its first variable instead of losing it, and the startup refusal for a byte-order mark added in 6.4.0 no longer fires on a pinned install (it still names the cause on an older python-dotenv).

### Upgrading

- **Run `pip install -r requirements.lock` after pulling** to move to python-dotenv 1.2.4.

## [6.4.0] — 2026-10-03

API keys can now be read from a JSON file, and a `.env` line python-dotenv cannot use stops startup with its line number instead of a one-line warning that is easy to miss.

### Added

- **`API_KEYS_FILE` reads the key array from a file (PR #142).** Set it to a path (recommended: `project_config/api_keys.json`, which is git-ignored) and put there exactly what `API_KEYS_JSON` holds, formatted however you like, with none of `.env`'s quoting rules. A relative path resolves against the repository root, like `PROJECT_CONFIG_DIR`. The file is read once per process, like an environment variable, so edit it and restart. Setting both `API_KEYS_FILE` and `API_KEYS_JSON` is refused. A missing or unreadable file, a file that is not UTF-8, and invalid JSON (the message gives the line and column) are refused naming the path, never the content. Messages and the "no `denied_columns` field" warning now name the source, for example `API_KEYS_FILE (project_config/api_keys.json)[0] (id='x') ...`, and `verify_deployment.py`, the startup refusal and `issue_api_key.py` mention the file option.

### Changed

- **Unusable `.env` lines are refused at startup with line numbers and names (PR #142).** python-dotenv skips a line it cannot parse with one line on stderr, so a pretty-printed `API_KEYS_JSON` without single quotes, or with an apostrophe in a value such as `"Ali's key"`, ended as "no usable key is configured". `Settings.validate()` now lists every unparsable line, every leftover line of a broken multi-line value (and the variable whose value was cut off at its first line break), and every variable assigned twice with different values (names and line numbers only, never values), and a byte-order mark at the start of the file that hides its first variable ("UTF-8 with BOM", which python-dotenv 1.2.2 as pinned in `requirements.lock` reads as part of the name; newer releases strip it) is named as such, so the server and the `Settings.validate()` check in `verify_deployment.py` fail with the cause. The same variable set twice to the same value is not flagged, and neither is a missing `.env`.
- **A repeated field inside one key object is refused (PR #142).** `json.loads` kept the last of two identical fields silently, so a pasted second `"denied_columns"` replaced the first. This applies to `API_KEYS_JSON` as well as the file. The startup refusal for a bad key configuration now begins `Invalid API key configuration:` instead of `Invalid API_KEYS_JSON:`, because the cause may be in the file.

### Upgrading

- **A `.env` that has an unparsable line, a leftover line, or a variable assigned twice with different values now stops startup.** Fix the lines the message names: wrap a multi-line `API_KEYS_JSON` in single quotes with no apostrophe inside, or delete the stale duplicate (the later one was the one in effect).
- **To move keys to a file:** create `project_config/api_keys.json` containing the array, set `API_KEYS_FILE=project_config/api_keys.json`, remove `API_KEYS_JSON`, restart, and run `python scripts/verify_deployment.py`.

## [6.3.2] — 2026-10-03

A key in `API_KEYS_JSON` with no `denied_columns` field is now warned about once, not every time the key list is read.

### Fixed

- **The "no `denied_columns` field" warning is logged once per key (PR #140).** The key list is read several times: at startup, by `python -m scripts.verify_deployment`, and again every `KEY_CACHE_TTL_SECONDS` while the server runs. Each read logged the warning again for every such key, so `verify_deployment` printed each one twice, and a running server repeated them at startup and on every key-cache refresh, burying real log lines. Each key is now warned about once per process; a new key, or the same id with a new hash, warns again, and adding `"denied_columns": []` stops the warning. Which keys get the warning, and what it says, are unchanged.

## [6.3.1] — 2026-10-03

A configuration file that repeats a key is now refused instead of silently losing the first one, and the deployment check counts API keys the way the server does.

### Fixed

- **A duplicate key in a YAML config file is refused (PR #138).** `yaml.safe_load` kept the last of two identical keys without a word, so a second `datasources:` heading dropped the first block, a whole data source, with no error. Every config file is now read so that a key written twice in one mapping fails with its name and both line numbers, for example `[datasources.yaml] is not valid YAML: duplicate key 'datasources' (first on line 12, again on line 21)`. An explicit key that overrides one brought in by a merge key (`<<: *anchor`) is still allowed.
- **`verify_deployment.py` counts application-database keys (PR #138).** The API key check counted only `API_KEYS_JSON`, so it failed on a deployment whose keys live in the application database (imported at first start, or issued from the admin panel) even though the server starts. It now counts the same merged set the server does, says how many keys came from each place, and resolves `VERIFY_API_KEY` against it. If the database cannot be read it counts `API_KEYS_JSON` only and says so; it fails only when that has no keys either.

### Upgrading

- **A config file that repeats a key is now refused,** naming the key and both lines. A duplicate in `datasources.yaml` or `schema.yaml` stops the server at startup; one in the other files fails when that file is first read, and `python scripts/verify_deployment.py` reports it before then. `relationships.yaml` is optional and is skipped with a warning naming the duplicate, as with any other read error. Remove the duplicate; the later one was the one in effect, so keep that block's content if it is what you meant.

## [6.3.0] — 2026-10-02

A warehouse connection is now described in `datasources.yaml`, and `.env`
holds only the raw password. Existing setups keep working unchanged.

### Added

- **Structured data sources (PR #136).** A source in `project_config/datasources.yaml` may give `host`, `port` (default 1433), `database`, `driver` (default `ODBC Driver 18 for SQL Server`), a login (`username`, or `username_env` naming a variable) and `password_env`, the name of the variable holding the **raw** password. `trusted_connection: true` selects Windows authentication and takes no login. Extra ODBC keywords go under `options`; YAML `true`/`false` become `yes`/`no` and numbers become their text.
  - The URL is built in one place with SQLAlchemy's `URL.create`, which escapes the password itself. A password containing `@ % ] : / ? # & = + ;` or spaces needs no encoding.
  - Two databases on one server are two sources, each with its own pool. One query still runs on one source; keep a single source with a multi-part `db_schema` when questions join across databases.
- **`DB_PASSWORD` (PR #136).** Without `datasources.yaml`, the raw password for `DB_CONNECTION_URL` can be set here instead of being percent-encoded inside the URL. It is applied to the parsed URL. A password in both places is refused at startup.
- **Clear refusals at startup (PR #136).** A `password_env` or `username_env` variable that is unset or empty is refused naming the source and the variable, never the value. A source that sets both `url_env` and structured fields, a `password:`/`pwd:`/`url:` key in the file, a credential-like `options` key (`pwd`, `password`, `uid`, `user`, `user id`) and a placeholder host, database, user name or password are refused too.

### Changed

- **`url_env` is the legacy form (PR #136).** It works exactly as in 6.1 and 6.2, may share a file with structured sources, and still needs the password inside the URL percent-encoded (`@` becomes `%40`).
- **`.env.example` and the example `datasources.example.yaml` (PR #136)** now show the structured form and `DB_PASSWORD_*` variables. The `DB_URL_*` examples are gone.

### Security

- **Passwords are masked in messages (PR #136).** The refusal when `APP_DB_URL` points at a warehouse source used to print both URLs in full, passwords included. It now masks them, as `scripts/verify_deployment.py` does for the connectivity check.
- **urllib3 2.8.0 (PR #136).** The pinned urllib3 in `requirements.lock` moves from 2.7.0 to 2.8.0, which fixes PYSEC-2026-4175, PYSEC-2026-4176 and PYSEC-2026-4177.

### Upgrading

- No action needed. A deployment without `datasources.yaml` is unchanged, and so is one that uses `url_env`.
- **To use `DB_PASSWORD`,** remove the password from `DB_CONNECTION_URL` and put it, raw, in `DB_PASSWORD`. A raw password written inside the URL is not detected or re-encoded, because `/ ? # :` or a literal `%40` cannot be told apart from an encoded one.
- **To move a `url_env` source to the structured form,** copy the host, port, database, login and query parameters from its URL into the YAML (query parameters such as `TrustServerCertificate=yes` become `options`), put only the raw password in a new `DB_PASSWORD_*` variable, replace `url_env` with `password_env`, restart, and run `python scripts/verify_deployment.py`. Remove the old `DB_URL_*` variable afterwards.

## [6.2.0] — 2026-09-28

**Issues fixed:** #130, first part (the same table name in two schemas: schema-qualified keys).

Tables with the same name in different schemas can be described and
queried, and a query's schema is now checked against the allowlist.

### Fixed

- **`schema.yaml` table keys may name their schema (PR #133).** A key may be qualified, such as `sales.Customer` or `OtherDb.dbo.Customer`; brackets are allowed. This is what lets `sales.Customer` and `ref.Customer` both be described. Before, a qualified key never matched a query, so every query on it was refused as an unknown table.
- **An ambiguous table name is refused with the fix named (PR #133).** `FROM Customer`, when two schemas have a `Customer`, is refused with the new guard reason `ambiguous_table`. The message names the references to use, so the model's retry corrects it, and the web UI explains it in Persian.
- **The prompt shows how to reference a shared name (PR #133).** Such tables get a `Reference as: [sales].[Customer]` line. Prompts for schemas with unique names are unchanged.

### Security

- **A query's schema is checked, not ignored (PR #133).** Before, the guard matched tables by bare name only, so with `Customer` allowlisted in `sales`, a query on `[hr].[Customer]` passed and read a table that is not allowlisted. An explicitly qualified reference must now match the table's known schema. A table with no known schema (a bare key without `db_schema`) still accepts any schema, because there is nothing to check it against.

### Upgrading

- No action needed for a `schema.yaml` with unique table names. A generated query that names the wrong schema is now refused, and the model's retry corrects it.
- **If the same table name exists in several schemas,** key each one with its schema, such as `sales.Customer:` and `ref.Customer:`.
- **Set `db_schema` on every bare-keyed table.** Without it, that table's schema cannot be checked. See `docs/design/TABLE-NAMES.md`.

## [6.1.0] — 2026-09-28

A deployment can query more than one warehouse database, including
databases on different servers. Existing setups keep working unchanged.

### Added

- **Multiple warehouse data sources (PR #131).** `project_config/datasources.yaml` is optional and lists named sources. Each source names only the environment variable holding its connection string, for example `url_env: DB_URL_MAIN`. The connection string itself stays in `.env`, because `project_config/` is versioned and reviewed.
  - Without the file there is one source, `default`, using `DB_CONNECTION_URL`, exactly as before.
  - A template is in `project_config.example/datasources.example.yaml`; the design and roadmap are in `docs/design/DATASOURCES.md`.
- **Each table names its source (PR #131).** A `schema.yaml` table may set `datasource:`; a table without it belongs to the default source. Several databases on one server are one source: a table in another database on that server sets a multi-part `db_schema: "OtherDb.dbo"`, rendered `[OtherDb].[dbo].[Table]`.
- **Queries are routed by their tables (PR #131).** The source is derived from the tables a query reads, never chosen by the model. One query runs on one source. A query whose tables span two sources is refused with the new guard reason `cross_datasource`, and the web UI explains why in Persian.
- **Every source is checked (PR #131).**
  - Startup refuses an application database that matches any source, and a table assigned to a source that is not configured. Admin-panel `schema.yaml` drafts are checked the same way.
  - `/health` pings every source; `database_detail` names each one when there are several.
  - `scripts/verify_deployment.py` and the admin panel run the connectivity, read-only login, row cap and query timeout checks once per source.
  - Schema drift compares each source's tables against that source's own server.
- **The audit record names the data source (PR #131).** The new `datasource` field is the source the query's SQL targeted. Older records simply lack it.

### Upgrading

- No action needed for a single-database setup.
- **To add a second database on the same server,** keep one source and give the table a multi-part `db_schema`, such as `OtherDb.dbo`.
- **To add a database on another server,** copy `datasources.example.yaml` to `project_config/datasources.yaml`, set one `DB_URL_*` variable per source in `.env`, add `datasource:` to that server's tables in `schema.yaml`, and restart. Create the read-only login on every source's server (`docs/db-hardening.md`), then run `python scripts/verify_deployment.py`.
- Changing `datasources.yaml` needs a restart.

## [6.0.2] — 2026-09-28

Running the API and the web UI on different ports or hosts no longer needs
guesswork.

### Added

- **`python -m api` starts the server from settings (PR #127).** It uses `API_HOST` (default `127.0.0.1`) and `API_PORT` (default `8000`) and sends no `Server` header. An invalid `API_PORT` fails with a message naming it. The documented `uvicorn api.server:app ...` command still works unchanged.
- **The server logs its allowed CORS origins at startup (PR #127).** It also shows a hint in both pages when the backend cannot be reached. The hint names the two likely causes (the backend is not running, or this page's origin is missing from `CORS_ALLOWED_ORIGINS`) and prints the origin to add.

### Fixed

- **The admin panel uses the same API address as the main UI (PR #127).** Both pages read it from `web/js/config.js`, with the same precedence: `?base=`, then a saved value, then the file. If `DEFAULT_BASE_URL` is empty, both derive the address from the page's own host and `DEFAULT_API_PORT`. Before this, the admin panel ignored the file and started from `http://localhost:8000`.
- **API addresses are normalised and validated (PR #127).** An address typed without `http://` is no longer treated as a relative path. Previously that sent requests such as `/admin/<host>:<port>/admin/...` to the static server. Now:
  - a missing scheme is added;
  - whitespace, paths and credentials are dropped;
  - IPv6 addresses work;
  - anything else is rejected with a message, and the previous address stays in use.
- **The application's own log lines are shown (PR #127).** INFO-level lines, such as the startup banner and the CORS line, were silently dropped under both documented start commands. When nothing else has configured logging, they now go to stderr at `LOG_LEVEL` (default `INFO`).

### Upgrading

- No action needed for an existing setup.
- **To use `python -m api` on a server reached from other machines,** set `API_HOST=0.0.0.0`; its default listens on the local machine only.
- **If the UI is served from anywhere other than `localhost:8080`,** set `CORS_ALLOWED_ORIGINS` to that origin, for example `http://172.16.101.42:8077`.

## [6.0.1] — 2026-09-27

The remaining `SELECT 1` traffic to the warehouse is gone. A diagnostic kit
helps a DBA find what is really driving disk activity.

### Changed

- **Pooled connections are pinged only after they have been idle (PR #124).** With `pool_pre_ping`, every checkout of a warehouse connection sent a `SELECT 1` first, so the warehouse saw one ping per query.
  - A connection is now pinged only if it has been idle for at least `DB_POOL_PING_IDLE_SECONDS` (default 60). `0` pings on every checkout, and `DB_POOL_PRE_PING=false` never pings.
  - A failed ping refreshes the whole pool, and the checkout is retried on a fresh connection.
  - Measured on 20 back-to-back questions: 20 pings in 5.2.0 and 6.0.0, none now. An idle application sends nothing.
  - Trade-off: a connection the server drops before it reaches the idle threshold makes one query fail before the pool recovers.
- **`/health` always sends its own `SELECT 1` (PR #124),** so it still verifies the database when the checkout was not pinged.
- **Comments and test docstrings describe behaviour, not history (PR #125).**

### Added

- **A read-only diagnostic kit for the DBA (PR #124).** `docs/dba/warehouse-load-diagnostics.sql` and `docs/dba/README.md` show:
  - whether `AUTO_CLOSE` is on;
  - this application's sessions (`program_name = 'local-sql-agent'`);
  - which database file (data, log or tempdb) is busy;
  - the queries with the most physical reads;
  - login audit and trigger checks.

  `SELECT 1` reads no data pages, so disk activity that coincides with it usually comes from something else.

### Upgrading

- No action needed. The new `DB_POOL_PING_IDLE_SECONDS` defaults to 60; set it to `0` to keep pinging on every checkout.
- Upgrading from 5.x: follow the **Upgrading** steps of 6.0.0 first (the system prompt must be at `<PROJECT_CONFIG_DIR>/system_prompt.md`).

## [6.0.0] — 2026-09-27

A release about trust in what the screen says, and about the load the
application puts on the systems around it.

A hall named in a question could be silently ignored, and a filter chip could
show a value the SQL never applied. Both now either work or say plainly that
they did not. Two concurrency defects are fixed: the request cap did not
cover the analyst's main route, and the SQLite application database shared a
single connection between every request thread. The admin panel stops
re-running expensive warehouse checks every thirty seconds.

The major version is for one operational change. The system prompt now lives
in the deployment's own project-config directory, and the server will not
start until it is there. See **Upgrading**.

### Fixed

- **A hall named in the question now filters the answer (PR #119).** The dimension vocabulary only matched a question that contained a stored value in full, so naming just a hall's distinguishing word matched nothing. A new token tier matches on a value's distinctive words when the question names the dimension; when two values tie, the analyst is asked which one, never guessed for. Also in this PR:
  - The vocabulary is now warmed at startup by default, so the first question after a restart is no longer answered as if the value had not been named.
  - A named dimension whose vocabulary could not be checked produces a Persian warning instead of silence.
  - After generation, every text filter the turn presents as applied is checked against the SQL. A missing one triggers one regeneration and, if it is still missing, a Persian warning naming it.
  - Editing an inherited filter on an "among those…" turn now rebuilds the query rather than changing only the chip.
  - Choosing an answer in a "which one?" clarification is now sent to the server in live mode instead of changing only the local chip.
- **Every pipeline route is under the concurrency cap, for the whole response (PR #117).** The cap covered only `POST /query`, and a streaming response released its slot as soon as its headers were sent. It is now pure ASGI middleware and covers `/query`, `/query/stream`, v2 turns (streaming or not) and the assumptions `PATCH`. The slot is held until the body is fully sent, and released on error or disconnect. All blocking pipeline work shares one worker-thread bound (`QUERY_THREAD_LIMIT`). The 503 `SERVER_OVERLOAD` response carries the request id again.
- **The SQLite application database no longer shares one connection across threads (PR #120).** Concurrent requests could end each other's transactions. File-backed SQLite now gives each request its own connection, with WAL journal mode and a busy timeout, and its `-wal`/`-shm` files get the same restricted permissions as the database.
- **Approving an access request is all-or-nothing (PR #115).** Approval runs in one transaction. A failure part-way leaves the request open with no key changed, so it can simply be retried. An approval and a denial of the same request can no longer both succeed.
- **No more backend English in the failure banners (PR #118).** The generic database message is no longer shown under the Persian lead sentence. A database-specific message appears as a labelled technical detail («پیام پایگاه داده:»), and the guard's English rule text stays in the SQL panel rather than in the banner.
- **Table names in the migration tool's reseed step are quoted (PR #114),** and the web UI tests' Node timeout is long enough for slow CI runners.

### Changed

- **Much less load on the warehouse (PR #121).** A DBA reported a steady stream of `SELECT 1` from the application. Most of it came from the admin panel, which re-ran the deployment checks and a full schema-drift scan of the warehouse catalogue every 30 seconds. The deployment checks include a rolled-back `CREATE TABLE` and a `WAITFOR` probe.
  - Those cards now load when the page opens and on their own refresh button, with a server-side cache (`ADMIN_EXPENSIVE_CACHE_TTL_SECONDS`, default 300).
  - The DDL and `WAITFOR` checks run from the panel only when explicitly requested.
  - `/health` caches its database ping (`HEALTH_CACHE_TTL_SECONDS`, default 15).
  - `pool_pre_ping` is configurable (`DB_POOL_PRE_PING`).
  - SQL Server connections identify themselves as `local-sql-agent` (`DB_APPLICATION_NAME`).
  - The deployment runbook lists everything the application sends to the warehouse, and how often.
- **The public repository no longer names one deployment's schema (PR #122).** Its schema qualifiers and table names are replaced by the example schema's own names everywhere.
  - The example configuration is now self-consistent: every relationship and every golden-set case refers to tables and columns the example schema defines, and the example `Order` has foreign keys to `Ring`, `Broker` and `Symbol`.
  - `database/relationship_map.py` honours `PROJECT_CONFIG_DIR`.
  - A test scans the contents and paths of every tracked file for the removed names; it holds them only as salted hashes.
- **The system prompt moved to the project-config directory (PR #122).** It is now read from `<PROJECT_CONFIG_DIR>/system_prompt.md`, with a generic example in `project_config.example/`. `prompts/few_shots.md` and `prompts/business_glossary.md`, which nothing read at runtime, are removed.

### Upgrading

- **Required:** before upgrading, copy your current `prompts/system_prompt.md` to `<PROJECT_CONFIG_DIR>/system_prompt.md`. Without it:
  - the API server refuses to start with `RuntimeError: System prompt not found: <path>`;
  - `app.py` exits with status 1;
  - the eval CLI raises `FileNotFoundError`;
  - the webapp raises the same `RuntimeError`.
- A **relative** `PROJECT_CONFIG_DIR` is now resolved against the repository root, not the working directory. A service started from another directory should check it, or use an absolute path.
- Startup now reads the prefetched dimensions' values from the warehouse. Set `DIMENSION_VOCABULARY_WARM_ON_STARTUP=false` if startup must not touch the warehouse; set `DIMENSION_VOCABULARY_TOKEN_FALLBACK_ENABLED=false` to keep whole-value matching only.
- New optional settings: `ADMIN_EXPENSIVE_CACHE_TTL_SECONDS`, `HEALTH_CACHE_TTL_SECONDS`, `DB_POOL_PRE_PING`, `DB_APPLICATION_NAME` (see `.env.example`).
- Tests that check a deployment's own values now read them from optional files under `<PROJECT_CONFIG_DIR>/_test_fixtures/` (see `project_config.example/_test_fixtures/README.md`) and skip without them.

## [5.2.0] — 2026-09-24

A security fix and the first of the refusal actions the design has always
specified but the product never offered.

The security fix is the same class as finding #11 of the 4.12.1 audit, this
time for the database rather than the model endpoint: when a query failed, the
raw database error reached the analyst, carrying the server's host, port and
instance, the database login name, the database name, and the full SQL with its
bound parameter values. No passwords and no data rows were exposed, but it was
a map of the infrastructure handed to anyone who could make a query fail.

The rest is about what an analyst can do when the answer is "no". A refusal
over a restricted column now offers a way to ask for that column; a refusal of
any kind now shows the statement that was refused; and every failure is
explained in Persian, in the analyst's terms, instead of in the backend's
English. The test matrix also grew to Windows and macOS, which surfaced a real
bug in how the router measured latency on Windows.

### Security

- **Database error text no longer leaks infrastructure detail (PR #109).** Before this change, when a query failed, `database/executor.py` wrapped the raw SQLAlchemy error and both the v2 conversation path and the v1 `/query` path showed it to the analyst. Depending on the database that exposed the server host, port and instance, the driver name and version, the database login name, the database name, and — for any execution error — the full SQL and the bound parameter values. It is the same class of problem as finding #11 in the 4.12.1 audit (the model endpoint), this time for the database. Now one shared classifier (`database/errors.py`) handles every path: connection and login failures become `DATABASE_UNAVAILABLE` and timeouts become `QUERY_TIMEOUT`, both with generic messages; a malformed query shows only the database's own sentence (for example `Invalid column name 'X'. (207)`). It fails closed: database text is shown only when the error is positively recognised as a statement error, and a final check replaces any text that still contains an IP address, a port, `user@host`, or driver/host/login wording. The full raw error is still written to the server log. Recognised across SQL Server, PostgreSQL, MySQL and SQLite.

### Added

- **"Request access" on a denied-column refusal (PR #110).** When the guard refuses a question because a column is restricted for the analyst's account, the refusal card now offers «درخواست دسترسی». A `security` admin reviews the request in the admin panel and approves or denies it (a denial needs a reason, which the analyst can see; the analyst sees their requests' status in the account menu). Approval removes that column from `denied_columns` on every live key the person holds; revoked keys are never touched. The column is taken from the server-side audit record, never from what the browser sends, and approval goes through the existing permission-changing code, so no new code path can widen access. A second request for the same column while one is open merges into it. New table `access_requests`.
- **"See the SQL" on a guard refusal (PR #105).** The statement the guard refused is now kept, as `guard.rejected_sql`, and shown behind a collapsed «دیدن SQL», clearly labelled as not run. It deliberately does not reuse `Turn.sql`, which means "the SQL that ran". The refused statement is now also kept in the audit record, so operators can see what was blocked.
- **The audit-log report shows what the guard refused (PR #106).** `scripts/analyze_audit_log.py` gains a section with refusals counted by reason and by subject (for example `SERVERPROPERTY ×2`). Verbatim refused statements appear only with `--include-examples`, so the default report stays safe to copy off the server.

### Changed

- **Every failure the analyst sees now has a Persian sentence (PR #107).** One sentence per error code and one per guard-refusal reason; the backend's English message no longer appears in those banners (it stays in the server log). Failures raised at the HTTP level, such as maintenance mode, now keep their real code instead of being reported as a network error. The out-of-scope message no longer names a specific deployment's domain.
- **CI runs on Windows and macOS as well as Linux (PRs #101, #104),** across Python 3.11, 3.12 and 3.13. Getting there fixed a real bug: router latency budgets did not work on Windows before Python 3.13, where `time.monotonic()` resolves to only about 15.6 ms, so any backend call faster than that measured as zero and a budget breach was never detected; latency is now measured with `time.perf_counter()`.

### Fixed

- **Identifiers containing `]` no longer break retrieval queries (PR #103).** The two query builders in `retrieval/` wrapped names in square brackets by hand without escaping; they now use sqlglot to quote identifiers.
- **The CI doctest step no longer opens real database connections (PR #102).** Doctests in `eval/runner.py` started background threads that tried to reach the configured warehouse host.

### Upgrading

- The new `access_requests` table is created automatically when the application starts, like the other application-database tables, so the default setup needs no manual step.
- Deployments that manage the application database schema with Alembic instead should run `alembic upgrade head` from the repository root; it applies migration `0004_access_requests`.

## [5.1.0] — 2026-09-20

A security release for the SQL guard. Tables and columns had long since moved
to an allowlist, but functions and session variables were still governed by a
denylist — three names and two identifier prefixes. That asymmetry left a
reconnaissance channel open, and an adversarial review of the guard found it.

The gap had two spellings. A query that names **no table** never engages the
table/column allowlist at all: it escapes the allowlist rather than tripping
it. And a query that *does* name a real, allowlisted table could still carry a
metadata function in its `SELECT` list — `SERVERPROPERTY`, `OBJECT_NAME`,
`COL_NAME` and their relatives produce an output column that is neither a
column reference nor a table reference, so nothing inspected it. The second
spelling is the reason this release is broader than the table-reference rule
first proposed: requiring a table closes only the first.

Closing it properly meant finishing the job the module started — moving the
last construct from a denylist to an allowlist, which also covers metadata
functions SQL Server ships in future rather than only the ones found.

### Security

- **The guard now restricts every accepted query to warehouse-reading shape**:
  allowlisted tables and columns, read through ordinary data functions, and
  nothing else. Three rules run at the end of `validate_sql`, after every
  existing rule, so a query that already violated an earlier rule still reports
  that rule's own more specific reason — the only verdicts that change are
  ones that were previously *accepted*.
  - Server, session and connection state is refused wherever it appears in the
    tree — projection, `WHERE`, `ORDER BY`, a CTE body, a scalar subquery — not
    only in the `SELECT` list. `GETDATE()`/`CURRENT_TIMESTAMP` is deliberately
    exempt: it reads the wall clock, not the server's identity.
  - Unrecognised function calls are refused unless named in a new allowlist,
    seeded with the measured set of legitimate T-SQL data functions that
    `sqlglot` has no typed class for. Everything an ordinary analytic query
    uses — aggregates, casts, date parts, window functions, `STRING_AGG`,
    `IIF` — parses to a typed node and never reaches this check.
  - A query must reference at least one non-CTE table.

### Added

- A `no_table_reference` value on the guard verdict's `reason`, for the
  table-less case. It is additive: clients already fall back to the generic
  refusal action for any reason they do not specially handle. The v2 contract
  document now lists it, and a test asserts every reason the guard can emit
  appears there — the registration was two places and silently drifted to a
  stale document, so it is now three places with a failing build behind it.

### Fixed

- The rejection reported for a state-reading construct named the `sqlglot`
  class (`PARAMETER`, `CURRENTUSER`) rather than the construct. It now names
  what the analyst wrote (`@@VERSION`, `CURRENT_USER`, `OBJECT_ID`), with any
  call arguments cut so a model-generated literal never reaches the structured
  `subject` field — which the contract documents as analyst-facing and never
  internal detail.
- `anyio` is bumped past CVE-2026-63374 and CVE-2026-64847, both published
  after the pin was last refreshed and caught by the `pip-audit` CI gate.

## [5.0.0] — 2026-09-16

The analyst UI redesign. Major, because it changes muscle memory: the topbar,
the anatomy of a turn, the composer's position, and the admin navigation all
move. Nothing about the query path or the SQL guard changes; this release is
about what the screen tells you and how much of it you have to take on trust.

The organising idea is that a conversational analytics UI is a claims
interface. Every answer asserts something about the warehouse, and the screen
either shows what that claim rests on or asks the analyst to assume it. So
what a turn had buried — the resolved question, the assumptions applied, the
stage the pipeline reached, why a refusal happened — is now visible before
the result rather than folded into a drawer.

### Changed — analyst-visible

- **The run-mode switch is gone from the topbar.** Run mode decided whether
  the numbers came from the warehouse or from built-in fixtures, and nothing
  else on screen distinguished the two, so one mis-click sat between an
  analyst and a screen of synthetic figures that look exactly like real ones.
  Simulated mode stays fully available at `?live=0`, where choosing it takes
  intent, and remains the automatic fallback when a backend cannot serve the
  v2 conversational path.
- **Turn anatomy is outcome-first.** The resolved question, assumption chips
  and clarifications render *before* the result node, never only inside the
  details drawer. Each failure mode — guard rejection, model unavailable,
  truncated output, forbidden SQL, execution error — renders its own anatomy:
  what happened, why, and a real next action. A guard refusal is resolved
  before a result card is considered, so a refusal can no longer appear as a
  zero-row result.
- **The composer sits below the transcript**, pinned to the viewport bottom.
- **Admin gains a health summary rail**, a sticky section jump nav, and a
  global refresh with a last-updated stamp; per-section refresh is demoted.

### Added

- **Design tokens in one file** (`web/styles/tokens.css`). Surfaces import
  them and never redefine a hex — light and dark redefine the *same* custom
  properties, which is what stops the two themes drifting apart.
- **An inline SVG icon set** (`web/js/icons.js`), built with DOM APIs so CSP
  `script-src 'self'` stays clean.
- **A user menu** carrying theme, language and the API key — identity moved
  out of the topbar's open surface.
- **Generated SQL is prettified and highlighted** like a code editor, with
  T-SQL rules for bracket identifiers and `N'…'` strings that stock Prism gets
  wrong. Highlighting is presentation-only: the copied text stays byte-exact.
- **A catalog-driven chart engine.** Chart form is selected from a closed
  vocabulary of analytical jobs rather than by an agent, and the boundary is
  specified and pinned by tests — a wrong chart form manufactures a false
  narrative, and that error class must stay deterministic and testable.

### Fixed

- Demo mode died at boot: a `const` the boot path read was declared below its
  use, so module evaluation threw and the whole UI never painted.
- All SQL highlighting was silently dead — the "already patched" marker was an
  enumerable key on the grammar, so Prism iterated it as a token and threw.
- A settled result could sit behind the sticky composer.
- Line and split-bar charts were offered for non-sequential data because a
  calendar word was matched as a substring rather than a whole token.
- The line chart drew its time axis right-to-left in an RTL document.
- The site's flagship hero query filtered on an alias that was never joined,
  so the landing page advertised SQL that cannot run.
- The marketing site fetched four font families from Google's CDN on every
  visit, leaking visitor IPs from the page that sells local-only analytics.

### Accessibility

- Light-palette contrast now clears WCAG 1.4.3: the brand teal measured
  3.74:1 and secondary text 4.40:1 against a 4.5:1 floor. Hue and saturation
  are unchanged; only lightness moved.
- Service health no longer rides on colour alone (WCAG 1.4.1) — state is
  carried in text and shape, so it survives both colour-vision deficiency and
  a monochrome screen.
- Interactive control edges reach 3:1 via a dedicated token; RTL layout and
  24×24 touch targets corrected across the chrome.
- Chrome ships no emoji: font-dependent colour glyphs render differently per
  OS, ignore `currentColor`, and carry no reliable accessible name.

### Internal

- `Turn` errors carry a `request_id` and a structured guard reason.
- The design policy, its invariants, and the chart-engine boundary are
  written down under `docs/design/` and pinned by tests rather than convention.
- Includes everything in 4.12.1.

## [4.12.1] — 2026-09-14

A security release. An independent audit (code review across MLSecOps, data,
API/web and infrastructure, plus a live penetration test against a running
server) found twenty issues, four of them High. Nineteen were code-fixable and
are addressed here; the twentieth is a repository-history matter left to the
owner (see **Not in this release**). Every fix carries a regression test under
`tests/security_audit/`, and the two live-confirmed leaks below were re-tested
against a running server after the fix, not only in unit tests.

### Security

- **Admin panel: stored XSS that escalated `operations` to `security` is
  closed** (High). A `principal_id` carrying a `"` was accepted, stored, and
  re-served into an HTML attribute where `escapeHtml`'s `textContent` idiom
  never encoded the quote. Broken in four independent places: the escaper now
  encodes `"`/`'`, rows are built with `createElement`/`dataset` rather than
  string interpolation, `principal_id` carries `pattern=^[A-Za-z0-9._-]+$`, and
  both static pages ship a `<meta>` CSP whose `script-src` excludes
  `'unsafe-inline'`.
- **`webapp/` no longer bypasses the per-principal column ACL** (High). The
  Flask app called the query pipeline with no principal, which the runner read
  as "no column restriction", so every `denied_columns` rule the admin panel
  could set was void on that path. It now maps the logged-in user to a real
  `Principal`, and an unmapped user gets a maximally restrictive fallback rather
  than unrestricted access.
- **Secrets and data files are created with owner-only permissions** (High).
  Nothing in the tree restricted file modes, so the Flask session-signing key,
  the app database, the session store and exports were world-readable on a
  POSIX host — a local shell could forge an admin session. A single
  `core/fileperms.py` authority now applies `0o600`/`0o700` at every creation
  point, SQLite `-wal`/`-shm` sidecars included.
- **The internal model endpoint no longer leaks in error text** (Medium; found
  Low, raised after a live test). A model-unreachable failure returned the
  backend's host, port, scheme and retry policy to any authenticated analyst.
  Both paths are fixed: the v1 `/query` path moves the raw transport error to a
  server-only `detail`, and the v2 conversational turn — a **second leak the
  first pass missed**, caught by a post-remediation live re-test — logs the raw
  error server-side and returns a generic summary, since `TurnErrorInfo` has no
  client-safe field to carry detail.
- **HTTP security headers are sent** (Medium). Responses now carry
  `X-Content-Type-Options`, `X-Frame-Options` and `Referrer-Policy`. The
  middleware also strips an app-set `Server` header, but uvicorn appends its own
  at the protocol layer where no middleware can reach it, so the documented
  start command now passes `--no-server-header` — the only place the banner is
  actually suppressed on the wire (`README.md`, `docs/deployment-runbook.md`).
- **`/health` is cached, so an unauthenticated flood cannot drain the query
  pool** (Medium). The probe result is cached briefly; a load balancer and an
  attacker both get the cached answer instead of one live pool probe per call.
- **CSV/Excel exports defuse formula injection** (Medium). A cell beginning
  `= + - @` (or a leading tab/CR) is prefixed so a spreadsheet cannot execute
  it, on both the API and Flask export paths.
- **Warehouse values are fenced as untrusted data in the model prompt**
  (Medium), reducing second-order prompt injection through dimension values
  pulled from the database.
- **Flask login hardened** (Medium): per-IP failure throttling with a window
  that lifts on its own (a throttle that never lifted was a denial of service in
  its own right), session cookies set `Secure`/`HttpOnly`/`SameSite=Strict`,
  and a session-bound CSRF token is enforced on every POST including `/login`.
  The real administrator username is no longer a baked-in default; a missing
  `ADMIN_USER` is now a startup failure rather than a guess.
- **Dependencies are pinned** with a `requirements.lock`, and CI runs
  `pip-audit` against it so a known-vulnerable pinned dependency fails the
  build.
- **Lower-severity hardening**: audit-log write failures surface a counter
  rather than failing silently; a dormant cache-scope parameter is documented
  as a trap before the future cache tier that would trip it; identifier quoting,
  the asymmetric default ACL between env and panel keys, and a global session
  cache with no per-owner quota are all addressed.

### Fixed

- A `frame-ancestors` directive that a `<meta>` tag cannot deliver was removed
  and the equivalent reverse-proxy configuration documented instead, so the
  page no longer logs a console error on every load.

## [4.12.0] — 2026-09-08

### Added

- **The conversational path can summarise its own result, and the analyst
  decides whether it does.** It never could: `session/engine.py` set
  `interpretation=None` as a literal from the commit that introduced the
  file, and never ran an `interpret` stage at all — so the web UI's fifth
  pipeline step, labelled "تفسیر", could only ever sit at "در انتظار".
  That was invisible while none of the five steps moved, and conspicuous
  the moment 4.11.0 made the other four tick.

  Producing a summary sends up to twenty rows of **real query results** to
  the model, so it is opt-in per request rather than on for a whole
  deployment: `AskTurnRequest.interpret` defaults to `false`, matching
  `/query`'s own default since phase 2, and the web UI exposes it as a
  checkbox under the question box that each analyst sets for themselves
  and that persists in their own browser. The label states the cost, not
  just the benefit — the person who asked the question is the one who
  knows whether these particular rows should go to a model.

  Asking for it also runs an `interpret` stage, so the fifth step reports
  what actually happened rather than staying decorative.

### Changed

- **The interpretation logic moved to `llm/interpret.py`.** It lived in
  `api/runner.py`, reachable only from `/query`; the session layer sits
  below the API layer and cannot import from it. Moved rather than copied,
  because it carries a data-governance gate — the refusal to send rows to
  a remote backend without `LLM_ALLOW_REMOTE` — and two copies of a gate
  is one copy that gets fixed and one that does not.
  `api.runner._interpret` is now a thin wrapper, and its existing gate
  tests cover the shared implementation unchanged.

---

## [4.11.2] — 2026-09-08

### Fixed

- **The admin panel laid out its English output right-to-left.** The page
  is `dir="rtl"`, which is right for its Persian chrome and wrong for what
  it displays: deployment-check cards carry
  `scripts/verify_deployment.py`'s own text — exceptions, SQL, connection
  URLs, Windows paths. Rendered RTL, `Settings.validate()` displayed as
  `()Settings.validate`, and a connection URL broke across lines in the
  wrong order. This is output an operator reads character by character to
  find a typo in. The check card is now LTR as a block, since its name,
  status pill and detail are one English sentence.

- **The keys table showed `.env` as `env.`** Its leading dot is a neutral
  character, so an RTL paragraph moved it to the end — a filename
  displayed wrong, in the column that says where a key came from. Cells in
  that table, and values in the key/value lists, now take direction from
  their own content (`unicode-bidi: plaintext`), because the same field
  legitimately holds a Persian maintenance note on one row and a path on
  the next.

- **Principal ids and role names are isolated**, so a hyphen inside
  `ops-1` cannot be reordered by the surrounding paragraph — the treatment
  `style.css` already gives numbers via `.num`.

### Added

- `tests/web_ui/test_web_ui_admin_direction.py`. The `.env` case is why
  these are assertions rather than a review habit: four characters, looks
  almost right, and invisible in the markup. It was found by rendering the
  panel and looking at it.

### Checked, and already correct

The rest of the panel was audited against the same failures found in the
analyst UI — page-level overflow, content clipped outside a scrollable
ancestor, labels rendered outside their container — at 1280px, 900px and
760px. None of them are present: the panel has no SVG, so the chart's
anchor bug does not apply, and the topbar inherits the wrapping fix from
4.11.1's shared `style.css`.

---

## [4.11.1] — 2026-09-08

### Fixed

- **Chart labels rendered outside the chart.** `web/js/render/chart.js`
  computes every position left-to-right — x grows rightward, and
  `edgeSafeLabel` flips `text-anchor` between `"start"` and `"end"` to keep
  a label inside `[0, W]`. But `text-anchor` resolves against the inline
  base direction, and both host pages are `dir="rtl"`, so the SVG
  inherited rtl and every one of those decisions meant its opposite.

  Measured in a browser: a label clamped to `x=2` with anchor `"start"`
  rendered at `x=-140`, 140 units outside the box — on the exact side the
  clamp existed to protect. Every label the function tried to rescue was
  precisely the one it threw out. The SVG now pins `direction: ltr`;
  Persian text inside is unaffected, since base direction sets where a run
  is anchored, not how it is shaped.

- **A mispositioned label escaped the chart entirely.** `.chart-block svg`
  carried `overflow: visible`, so a label outside the viewBox drew over
  the card around it rather than being clipped. Now `hidden`: a future
  geometry mistake reads as a cut-off label inside the chart instead of
  text across the layout.

- **Summed rial figures overflowed their summary tile.** The strip's grid
  tracks were sized at 140px for counts and percentages; a twenty-digit
  total at 17px needs far more, and a grid item's default `min-width: auto`
  meant the `1fr` track could not shrink to contain it. Tracks widened,
  `min-width: 0` so the track wins, and the value now scales down and
  wraps before it would clip — the digits cut in RTL are the most
  significant ones.

- **Ranking labels truncated to about twelve characters**, which turns
  most real company names into an ellipsis and a hint. The category track
  goes from 130px to 176px; the bars lose the difference and stay legible,
  where an unreadable label makes the whole ranking unreadable.

### Added

- `tests/web_ui/test_web_ui_chart_direction.py`. The existing chart test
  drives the real module against a stub DOM with no layout, fonts or bidi
  algorithm, so this class of bug was structurally invisible to it. The
  new test asserts the declaration whose absence *was* the defect, and
  checks its own premise: if `chart.js` ever resolves anchors against the
  document direction, the test says to delete itself along with the CSS
  rather than work around it.

---

## [4.11.0] — 2026-09-08

### Fixed

- **The web UI's five pipeline steps now actually tick.** They never did.
  The SSE endpoint emitted exactly one `stage` event, before the engine
  was even called, naming a stage `"plan"` — an id no step has. So
  `renderPipeline`'s `setStage` looked for `[data-step="plan"]`, found
  nothing, and returned without a word; all five steps sat at "در انتظار"
  for the whole turn while the answer appeared beside them.

  Nothing in the stack was broken enough to notice: the event was
  well-formed, the stream was valid, the turn was correct, and the one
  test covering the stream asserted the single frame as if it were the
  contract.

  Progress is now live. `StageTimer` already knew when each stage started
  and finished and had no way to say so; it takes an optional observer,
  `TurnEngine.ask` threads one through, and the endpoint bridges the
  worker thread back to the event stream with a queue. Each of the five
  steps reports `running` when it starts and `done` — or `error` — when it
  ends, including the repeats a correction round genuinely produces.

- **Generated SQL is formatted before it is displayed.** `sql_display` was
  the model's own text, so its layout was the model's mood: a tidy
  multi-line statement for one question and a single 300-character line
  for the next, which reads as the system formatting sometimes and not
  others. It is now re-rendered with sqlglot — already a dependency, and
  already the thing that parsed this SQL to validate it.

  Display only. `Turn.sql` stays byte-for-byte what the guard validated
  and the database ran; re-rendering that would mean the audit trail
  recorded a statement nobody executed. `pretty_sql` never raises — SQL it
  cannot re-render is shown as generated, because no formatting problem is
  worth failing a successful query over.

---

## [4.10.3] — 2026-09-08

### Added

- **Three badges for guarantees a build step actually enforces**, each
  linking to the step that makes it true: `SQL guard: AST allowlist`
  (`security/sql_guard.py`, parsing with sqlglot rather than matching
  strings), `engine: no domain literals`
  (`tests/test_no_domain_literals.py`, which fails the build if a real
  warehouse name reappears in engine source), and
  `doctests: enforced in CI` (a required workflow step across fourteen
  packages, not an optional one).

  `tests/test_readme_claims.py` covers these the way it already covers the
  coverage figure, but against a different failure: not drift, but the
  backing quietly going away. Each badge must still be present, still link
  to its evidence, and that evidence must still exist and still contain
  the thing being claimed — so a guard deleted in a refactor takes the
  badge down with it instead of leaving the README promising something
  that stopped being true.

---

## [4.10.2] — 2026-09-08

### Changed

- **Coverage gate raised 85% → 90%.** The README now advertises the
  measured figure, and a badge saying 92% with a gate at 85 is a number
  nobody is holding: coverage could fall seven points with every build
  still green and the README still claiming otherwise. Measured coverage
  is 92.84%, so 90 leaves real headroom without letting the claim rot.

- **README badges rebuilt around what can be verified.** The CI status
  and the releas