[0x054]

Panel round 1 on #777, the harness and page findings (the src/ findings landed in #779; this branch is rebased on it).

Part of #698 (panel round 1 on PR #777)

What a reader gets

Before, the output under "It prints:" was typed by hand. The aggregations page showed

TypeError: t.amount.sum() > ... is not a where() predicate: ... Post-aggregation filtering is having() ...

and nothing checked it. Now every ```text block on a page carries a note naming where it comes from, and the test regenerates it:

<!-- prints: aggregations.py:error-where -->
```text
TypeError: t.amount.sum() > ... is not a where() predicate: WHERE filters rows before aggregation, so an aggregate cannot appear in it. Post-aggregation filtering is having() (#291); until it lands, filter rows with where() and compare the aggregated result in Python.

The Annotated tab now shows the same program as the Assignment tab. Before, `mutations_annotated.py` was a different, shorter script. Now each twin is generated from its Assignment script, and a test fails when the two differ in anything but field syntax.

## Per finding

### Spec: `ferro migrate up --allow-ahead` missing from the CLI reference (Important)
- Changed: `reference/cli.md` lists `--allow-ahead=expand-only` and `--allow-ahead`. After the rebase it also lists `status --no-steps` / `--no-json`, which #779 restored; the new test caught both.
- Test: `tests/test_docs_cli_reference.py` (F24 below).

### Spec: the RLS how-to sat in the Guide tab (Minor)
- Changed: the nav entry moved to the How-to tab. The page keeps its path `guide/row-level-security.md`, because two `src/` docstrings link it (`models.py:278`, `rowsecurity.py:283`). Every link still resolves and no redirect is needed. Moving the file is a follow-up for whoever next touches those docstrings.
- Test: `test_the_row_level_security_split_sits_in_its_three_tabs` (tests/test_docs_site.py).

### F7: printed-output blocks never compared (Important)
- Changed: the marker is an HTML comment on the line above the fence, the form the migrations pages already use (`<!-- output: id -->`):
  - `<!-- prints: <example>.py:<snippet> -->`: what that region prints.
  - `<!-- prints: above -->`: what the run fence above prints.
  - `<!-- shows: <what it is> -->`: text no script prints.
- All 116 text blocks are marked. 21 are regenerated from example regions and 3 from the run fences above them. The rest are migrate CLI output (already gated) or `shows:`.
- The six aggregation error regions now print the error they catch, so the page shows each message whole. `concepts.py` gained an `unsupported-scheme` region for the backends page's `InterfaceError`.
- Test: `tests/test_docs_output.py`. It refuses an unmarked block and compares byte for byte, naming the page and line.
- RED: `46 failed, 68 passed`. GREEN: `116 passed`.

### F8: the nine story-72 entry points (Important)
- Test: `test_each_entry_point_keeps_a_running_example`, plus a scratch-package test that names a compile-only entry point.

### F9: five fence parsers (Important)
- Changed: `tests._docs_harness.fenced_blocks(path, text, language=None)` returns a `Fence` (language, info, dedented body, opening line, indent, the note above it) and refuses an unclosed fence. `fenced_examples` uses it, as do the SQL, step-file, output-block, README and TOML readers.
- Test: `tests/test_docs_harness.py`.

### F10: five subprocess runners (Important)
- Changed: `run_python(args, *, timeout, env, cwd, input)` raises with the command, the exit code and both streams. It is used by `run_fence`, the script and RLS runners, the README test, `_run_example`, the autogenerate test, the fixtures' pytest run and the SQL capture.

### F11: twins not checked (Important)
- Changed: `tests/_docs_twins.py` rewrites an Assignment script into the Annotated idiom and leaves every other line alone. All 29 twins are regenerated with it.
- Test: `tests/test_docs_twins.py`. It fails when a twin is not its script rewritten (`python -m tests._docs_twins docs/examples/<name>.py` rewrites it), or when the pair declares a different SchemaIR or prints different stdout.
- RED: `40 failed, 22 passed`. GREEN: `63 passed`.

### F13: aggregations' first block (Important)
- Changed: the `global` region starts at the imports and `connect()`. The lead-in names all five aggregates, and the result sentence says five numbers.
- Test: the SQL and output gates cover the region.

### F14: SQLite caveat (Minor)
- Changed: one sentence in auto-migrate.md's partial-failure section.

### F15: `default_factory` inside `Annotated` (Minor)
- Changed: decided in docs-page-conventions.md. The factory stays an assigned `Field(default_factory=...)`; only the other options move inside `Annotated`. The twin generator applies it.
- Tested: `ty check` on a scratch pair reports `missing-argument` for the factory inside `Annotated` and nothing for the assigned form.
- Not added: a `ty` gate. `ty` is only in the `dev` group, not `ci-test`.
- Test: `test_a_default_factory_stays_an_assigned_field`.

### F17: CONTRIBUTING.md fence escaped every gate (Minor)
- Changed: `page_fences()` also reads Markdown that a page pulls in with `--8<--`. The fence is marked `{test=compile reason=...}`, because it is a test for this repository's suite and needs that suite's `db_url` fixture.

### F18: `MOVED_ANCHORS` overclaimed (Minor)
- Measured: of the h2–h4 anchors at f3e46f8, 170 have no target now.
- Changed: the six named anchors redirect. The docstrings now say what the map covers: the split pages, the deleted Next steps page and those six.

### F19: hand-listed CI docs tests (Minor)
- Changed: the CI job runs `tests/test_doc*.py`.
- Test: the pin requires that glob, and fails on any test module that imports the docs helpers without matching it.

### F21 / F22 / F23 (Minor)
- F22: capture, layout, markers and block checks moved to `tests/_docs_sql.py` (`captured(example, dialect, url)`, `check_block`, `check_same_on_postgres`, `check_printed`). One run records both the SQL and the printed lines of every region.
- F21: `check_fence(example, *, eval_example=None)` holds a page fence and a docstring fence to one contract.
- F23: the capture tests assert through `by_region()` and through `captured()` on a scratch example.

### F24: no `--help`-vs-reference test (Minor)
- Test: `tests/test_docs_cli_reference.py`, parametrized over the `migrate` verbs.

### F25: Postgres-only compile blocks (Minor), done in review round 1 (see the fix report below)
- Not a small harness addition. The blocks connect to literal URLs, so the runner would have to patch `connect()`. The fence parametrization would need a `postgres_only` split, and the RLS blocks need Litestar and FastAPI, which no test group installs. The blocks stay compile-only with their reasons.

### Coherence (Minor)
- Changed: an unmarked docstring fence is refused, as a page fence is. It passes now that #779 marked the 19 fences.
- Changed: the unused `aggregations` prelude is deleted. `test_every_prelude_is_cited_by_a_fence` keeps that true.

## Interfaces changed
- Page marker: `<!-- prints: <example>.py:<snippet> -->`, `<!-- prints: above -->` and `<!-- shows: <what it is> -->` above a ```text fence. A ```text block with none fails.
- `tests/_docs_harness.py`:
  - `Fence`
  - `fenced_blocks(path, text, language=None, *, line_offset=0)`
  - `page_fences(language=None)`
  - `page_examples()`
  - `run_python(args, *, timeout=180, env=None, cwd=REPO_ROOT, input=None, clean_env=False)`
  - `fence_mode(example)` (was `page_fence_mode`)
  - `check_fence(example, *, eval_example=None)`
- `tests/_docs_sql.py`:
  - `captured(example, dialect="sqlite", url=None) -> Run(regions, stdout, schemas)`
  - `Region(sent, printed)`
  - `check_block`, `check_same_on_postgres`, `check_printed`
  - entry point `python -m tests._docs_sql <example> <dialect> <out.json> [url]`
- `tests/_docs_twins.py`: `annotated_twin(source)`, `twin_pairs()`, and `python -m tests._docs_twins <script>...`.

## For the reviewer
- The regenerated aggregation errors show Ferro's real messages, which carry issue numbers (`(#291)`, `(#294)`). The page used to hide them behind `...`. Removing them is a `src/` change.
- A model Ferro refuses at class definition (two primary keys) stays in the registry, so `compile_registry_schema_ir()` raises afterwards. The capture records the refusal as that run's schema. This looks like a `src/` bug worth a ticket.
- (Superseded by the fix report below: the eleven error and warning blocks are now `prints:`.) 23 `text` blocks were `shows:`. Seven are a sketch, a version number, a file tree, a pytest session and a side-by-side layout. Five are agent prompts. The other eleven are errors and warnings from scenarios no example runs: an auto-migrate pass over an existing table, a held run lock, a database behind or ahead of its migrations. These are the next candidates for `prints:`.

## Verify
On the rebased head 0860323 (base c80e3eb):

```text
$ uv run --no-sync zensical build --clean --strict
No issues found
$ uv run --no-sync pytest tests/test_docs_examples.py tests/test_docs_sql.py tests/test_docs_site.py tests/test_docs_lint.py tests/test_docstring_examples.py tests/test_docs_migrations.py tests/test_docs_testing_fixtures.py tests/test_docs_output.py tests/test_docs_twins.py tests/test_docs_harness.py tests/test_docs_cli_reference.py -q --no-cov -p no:cacheprovider -n 8
916 passed, 109 skipped
$ FERRO_POSTGRES_URL=postgresql://postgres:password@localhost:5432/postgres uv run --no-sync pytest <same> --db-backends=sqlite,postgres
1025 passed
$ uv run --no-sync pytest --no-cov -p no:cacheprovider tests/test_doc*.py -n 8   # the CI job's form
932 passed, 109 skipped
$ uv run --no-sync python -m tests._docs_checks docs/pages docs/examples
exit 0
$ uv run --no-sync ruff check tests docs/examples && uv run --no-sync ruff format --check .
All checks passed! / 348 files already formatted

Verify block corrected: no

• Issue status: Docs rewrite: every snippet runs, every fact matches the code, and the API reference renders cleanly #698 stays open (this PR is part of the epic, not its close); no sub-issue to close.

Fix report: review round 1 (Spec FAIL on F7 and F25)

F7: the 12 error and warning blocks are now regenerated

Each block that quoted Ferro's own error or warning text is now a prints: block. An example region triggers the error and prints it:

Block	Region
auto-migrate.md: required column with no default	auto_migrate_errors.py:required-column
auto-migrate.md: SQLite check	auto_migrate_errors.py:sqlite-check
auto-migrate.md: kept foreign default	auto_migrate_errors.py:foreign-default (postgres)
auto-migrate.md: run lock, waiting	auto_migrate_errors.py:lock-wait. The example holds the SQLite lock file itself.
auto-migrate.md and configuration.md: run lock, gave up	auto_migrate_errors.py:lock-timeout, with lock_timeout = "1s"
deploying.md: behind its migrations	migration_errors.py:behind
deploying.md: ahead of its checkout	migration_errors.py:ahead
deploying.md: ahead by a contract step	migration_errors.py:ahead-by-contract
overview.md: tracked database	migration_errors.py:tracked
renames-and-retirements.md: retired label	migrations_retire_label.py:retired-write
reference/row-level-security.md: forced table	row_level_security.py:forced_warning (postgres)

Regenerating found three blocks that no longer matched Ferro:

• The forced-table warning also names 'document'.
• The warnings start with ferro auto-migrate: .
• The lock error names the SQLite schema main.

The page prose now says the example sets "1s" on SQLite. deploying.md's migration names follow the scenario's numbering (0002_author_slug, 0003_drop_bio), as does the API page's contract_steps example.

New form: <!-- prints: <example>.py:<snippet> (postgres) -->. The region runs on PostgreSQL as <example>.py <database-url> <role> under postgres_only. captured() gained args for this.

shows: is narrowed to text Ferro does not print: a sketch, a file tree, another tool's shell session, a prompt to paste. test_a_shows_block_holds_nothing_ferro_prints fails any shows: block that reads like Ferro output (ferro., ferro auto-migrate:, SomethingError:, N validation error for, connect().

• One exception, listed in VARIES with its reason: connections.md's two sample statement-log lines. Each ends with how long the statement took, which differs on every run, so no capture reproduces it byte for byte.

TDD

• RED: the tripwire test failed on 11 shows: blocks.
• GREEN:
  • tests/test_docs_output.py: 127 passed, 2 skipped on SQLite; 129 passed with --db-backends=sqlite,postgres.
  • tests/test_docs_twins.py stays green with the regenerated migrations_retire_label and row_level_security twins.

F25: backend=postgres

• New setting {test=run prelude=<name> backend=postgres}:
  • With Postgres, the fence runs under postgres_only. The prelude runner sends every connect() to the test database, whatever URL the prelude or block names.
  • Without Postgres, the fence only compiles, and its Postgres test shows as skipped.
• backend= on a mode other than run, or a value other than postgres, is refused.
• Converted 8 page blocks: raw-sql.md (3), connections.md (3), backends.md (1), auto-migrate.md (the label-addition connect).
• test_docstring_example_on_postgres runs docstring fences the same way.
• Note for the src owner: session.py:295 and :422 can drop test=compile for test=run prelude=<name> backend=postgres; the test that runs them is already in place.
• Left compile-only, with their reasons:
  • the four Litestar, FastAPI and ASGI middleware blocks on the RLS how-to (Litestar and FastAPI are not installed in any test group);
  • its JWT-claims block, which needs the auth_uid() function set up earlier on that page;
  • auto-migrate.md's two model-only tabs and its fails-partway block. The reviewer did not list these.
• Tests: test_docs_examples_on_postgres (8 blocks, all passing on Postgres), test_a_postgres_fence_runs_on_postgres, test_backend_needs_test_run, test_only_postgres_is_a_backend, test_a_postgres_fence_without_a_database_is_refused_by_run_fence.

Verify on 21ed328 (base c80e3eb)

$ uv run --no-sync zensical build --clean --strict
No issues found
$ uv run --no-sync pytest tests/test_doc*.py tests/test_docstring_examples.py -q --no-cov -p no:cacheprovider -n 8
948 passed, 121 skipped
$ FERRO_POSTGRES_URL=postgresql://postgres:password@localhost:5432/postgres uv run --no-sync pytest tests/test_doc*.py tests/test_docstring_examples.py ... --db-backends=sqlite,postgres
1070 passed, 1 skipped   (the skip: the Postgres docstring test, which no fence uses yet)
$ uv run --no-sync python -m tests._docs_checks docs/pages docs/examples
exit 0
$ uv run --no-sync ruff check tests docs/examples && uv run --no-sync ruff format --check .
All checks passed! / 350 files already formatted

Interfaces changed in this round:

• prints: takes (postgres).
• captured(..., args=()).
• check_printed(..., dialect="sqlite").
• run_fence(example, *, database_url=None) and check_fence(..., database_url=None).
• The fence setting backend=postgres.
• The prelude runner reads FERRO_DOCS_DATABASE_URL.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Hb9U9A7Zu9ypdHTsDvdScP

[0x054]

Spec: FAIL

Spec = the panel findings this slice owns (PR #777, comment 6319609050). No Closes #N: accepted, this is a fix slice under epic #698 with no sub-issue.

• Important · docs/pages/guide/schema/auto-migrate.md:67,91,190,236,243; deploying.md:68,129,144; overview.md:45; renames-and-retirements.md:168; reference/configuration.md:90; reference/row-level-security.md:206; docs/solutions/patterns/docs-page-conventions.md:95 at 0860323 · Missing · F7 is narrowed, not fixed. These 12 text blocks quote Ferro's own errors and warnings (the NOT NULL refusal, the SQLite check warning, the kept-default warning, both run-lock texts, PendingMigrationsError, both DatabaseAheadError texts, the tracked-database refusal, the retired-label ValidationError, the forced-RLS warning). They are marked <!-- shows: ... -->, which check_printed (tests/_docs_sql.py:583) never compares. The conventions now define shows: to include "an error from a scenario no example runs", so the exemption is part of the rule. Every text matches its source in src/ and crates/ today, so nothing is wrong yet. But a reworded message leaves a stale page with every gate green, which is the defect F7 names. These are also the texts most likely to change, and most of them can be produced cheaply: the retired-label error and the tracked-database refusal need only a model and a connect, and the deploying.md errors fit test_docs_migrations' existing refusal: scenario mechanism. There is no **Narrowed:** entry, and the PR body calls them "the next candidates for prints:" (I-6: a documented residual is not a resolution). · F7: "mark each output block with its example and snippet …, run the script once, record stdout per snippet region, and compare."
• Minor · docs/pages/guide/raw-sql.md:69,98,130; guide/connections.md:57,206,298; concepts/backends.md:88; guide/schema/auto-migrate.md:147; src/ferro/session.py:295,422 at 0860323 · Missing · F25 was not done, and the stated reason holds for only some of the blocks. The Litestar/FastAPI/ASGI reason covers the five row-level-security.md blocks. For these ten plain-Postgres blocks the other obstacle ("the runner would have to patch connect()") is one this PR already solves: _run_capture redirects ferro.connect to a given URL (tests/_docs_sql.py:374), and test_docs_sql already runs regions on Postgres. · F25: "these could run under a postgres-only fence setting."

Checked and addressed: Spec/story 17 (--allow-ahead both forms, cli.md:202-203; test_docs_cli_reference.py fails on c80e3eb's cli.md); Spec/RLS tab (zensical.toml:49, test_the_row_level_security_split_sits_in_its_three_tabs; the page keeps guide/row-level-security.md, and the links at src/ferro/models.py:278 and rowsecurity.py:283 resolve under --strict); F8 (test_docstring_examples.py RUNNING_ENTRY_POINTS, plus a scratch-package negative); F9 (fenced_blocks, _docs_harness.py:149; no other fence regex is left in the docs tests); F10 (run_python, _docs_harness.py:65; the remaining subprocess.run calls assert on non-zero exits and are outside F10's five); F11 (tests/_docs_twins.py: an AST rewrite plus ruff format, deterministic against the locked ruff 0.15.1, which pytest-examples pulls into ci-test; test_docs_twins.py compares source after conversion, SchemaIR and stdout for all 29 pairs); F13 (aggregations.py global region now mirrors predicates.py first-query; aggregations.md:9,15 name five aggregates and five numbers); F14 (auto-migrate.md:230); F15 (conventions lines added; no twin or page puts default_factory inside Annotated, e.g. timestamps_annotated.py:40); F17 (CONTRIBUTING.md:317 now collected, id CONTRIBUTING.md:317-334); F18 (six redirects in zensical.toml:195-200, each target id present in site/, docstring narrowed at test_docs_site.py:572); F19 (ci.yml:108 glob, plus the misnamed-module pin); F21 (check_fence); F22 (tests/_docs_sql.py); F23 (capture tests go through captured()/by_region()); F24; Coherence (unmarked docstring fence refused; prelude deleted, test_every_prelude_is_cited_by_a_fence).
Harness refactor, collected ids at c80e3eb vs 0860323: every existing parametrized gate collects the same blocks (only line numbers shift where notes were added), apart from one extra test_docs_examples id, the CONTRIBUTING fence (F17). test_docs_output collects 107 page blocks.
Cannot verify: whether the six redirected anchors are the ones outside links actually use (the docstring claims that).

Quality: PASS

• Minor · tests/_docs_sql.py:583 at 0860323 · check_printed passes any <!-- output: ... --> / <!-- refusal: ... --> block on any page. test_docs_migrations reads output: only on OUTPUT_PAGES and refusal: only on cli.md, so either note on another page escapes both gates. There is none today. · Fix: refuse those kinds outside the pages test_docs_migrations reads.
• Minor · tests/test_docs_site.py:316 at 0860323 · The new "docstring" sits after the first statement of test_ci_docs_job_runs_every_docs_test_and_the_lint, so it is a no-op string expression, not a docstring. · Fix: move it above jobs = ....
• Minor · PR body · The counts do not match the tree: there are 107 text blocks, not 116 (116 is the module's test count). 20 prints: blocks come from regions, not 21. 12 shows: blocks are errors or warnings, not 11, and 6 are other kinds, not 7.
• Depth check (_docs_sql, _docs_twins, fenced_blocks / run_python / check_fence): no check fired.

Verify: PR head 0860323 passed (zensical build --clean --strict "No issues found"; pytest tests/test_doc*.py tests/test_docstring_examples.py -n 8 932 passed, 109 skipped; the same with FERRO_POSTGRES_URL=… --db-backends=sqlite,postgres 1043 passed, Postgres accepted connections; python -m tests._docs_checks docs/pages docs/examples exit 0; ruff check tests docs/examples clean; ruff format --check . 348 files formatted). The capture of mutations.py is byte-identical under PYTHONHASHSEED 1, 7, 42 and 1234.
Trial merge: not run (base unchanged; merge-base = origin/feat/docs-rewrite = c80e3eb)

[0x054]

Scoped re-review: 0860323..21ed328

Spec: PASS

• F7 (Important, narrowed): ADDRESSED. The 12 Ferro error and warning blocks are now prints: blocks, each regenerated from a region: docs/examples/auto_migrate_errors.py (required-column, sqlite-check, lock-wait, lock-timeout, foreign-default), docs/examples/migration_errors.py (behind, ahead, ahead-by-contract, tracked), migrations_retire_label.py:77 (retired-write) and row_level_security.py:202 (forced_warning). The two Postgres ones carry (postgres) and run in test_a_postgres_text_block_is_what_its_example_prints_there (tests/test_docs_output.py:55), which CI's backend-matrix job reaches through -m "backend_matrix or postgres_only" (ci.yml:296). Regeneration corrected three stale texts: the ferro auto-migrate: prefix, 'document', 'invoice', and schema main with 1s. shows: is narrowed in docs-page-conventions.md:96 and guarded by test_a_shows_block_holds_nothing_ferro_prints (tests/test_docs_output.py:83). The renumbered migrations on deploying.md:124-152 and api/migrations.md:65-71 (0002_author_slug, 0003_author_slug_required, 0003_drop_bio) agree with each other and with what migration_errors.py prints.
• F25 (Minor): ADDRESSED. 8 page blocks are now {test=run prelude=… backend=postgres}: raw-sql.md:69,98,130, connections.md:57,206,298, backends.md:88, auto-migrate.md:147. All 8 are collected as test_docs_examples_on_postgres under postgres_only. Without Postgres they are reported as skipped, not passed, and the SQLite gate only compiles them (check_fence returns before run_fence, tests/_docs_harness.py:426). run_fence raises if asked to run one with no URL (pinned by test_a_postgres_fence_without_a_database_is_refused_by_run_fence). The 5 RLS blocks that stay compile-only (row-level-security.md:128,163,214,237,313) need Litestar, FastAPI, an ASGI server or the auth_uid() setup, which no test group installs, so that reason holds. session.py:295,422 are out of this slice.
  Cannot verify: none

Quality: PASS

• Minor · tests/test_docs_output.py:75-80 at 21ed328 · The one exception (VARIES) is honest about why: the statement-log lines end in a duration (src/ferro/_sql_log.py:54, {duration_ms:.1f}ms), and the block elides the column list. But it is not the I-6 answer, and the repo already has the primitive that is: test_docs_migrations' _normalise masks (\d+ ms) on both sides. A region that turns on the ferro.sql logger, plus a mask for the trailing duration (and a printed statement short enough not to need (...)), would make the block a prints: block. The exemption is also keyed by page, so any later shows: block on connections.md skips the check too. · Fix: regenerate with the duration masked, or at least key VARIES by page and line.
• Minor · tests/test_docs_output.py:68-71 at 21ed328 · FERRO_OUTPUT keys on a prefix (ferro., ferro auto-migrate:, …Error:, N validation error, connect(). Three of the twelve texts this round converted carried none of those prefixes when they were shows: (Cannot add NOT NULL column…, Check constraint 'ck_…, author.slug has default…), so the guard would not have caught them. The conventions sentence ("A shows: block that reads like Ferro's output … fails the test") claims more than the test enforces.
• Minor · docs/pages/concepts/backends.md:88, guide/connections.md:57,298 at 21ed328 · point_connect_at (docs/examples/_preludes/init.py:87) replaces the whole URL, so ?ferro_search_path=test_42, ?ferro_search_path=tenant_a and ?sslmode=require never reach connect(). Those blocks run, but the option each one demonstrates goes untested. auto-migrate.md:147 likewise connects to an empty database, so no label is added. · Fix (follow-up): carry the block's query parameters onto the test URL (a per-test schema name for ferro_search_path; drop only sslmode).
• Minor · docs/examples/migration_errors.py:63,77,90 at 21ed328 · The regions hand-write the exception's dotted name (f"ferro.migrations.DatabaseAheadError: {error}"), so a renamed or moved class would leave the page's first word stale with the gate green. · Fix: f"{type(error).__module__}.{type(error).__qualname__}: {error}", or the convention's type(exc).__name__.
• No new Critical or Important breakage. Both new scripts run under test_docs_example_scripts on SQLite. migration_errors.py works in project()'s temporary directory and leaves nothing behind. auto_migrate_errors.py's Postgres half runs only when given a URL.

Out-of-scope observations: none.

Verify: PR head 21ed328 passed (zensical build --clean --strict "No issues found"; pytest tests/test_doc*.py tests/test_docstring_examples.py -n 8 948 passed, 121 skipped, each Postgres test skipped with "No Postgres provider is configured"; with FERRO_POSTGRES_URL=… --db-backends=sqlite,postgres 1070 passed, 1 skipped (the empty docstring backend=postgres set), Postgres accepted connections; python -m tests._docs_checks docs/pages docs/examples exit 0; ruff check tests docs/examples clean; ruff format --check . 350 files formatted)
Trial merge: not run (base unchanged at c80e3eb)