AGENTS.md: rewrite I-8 for the lambda-only query API, add a docstring invariant, record the docs conventions #703

Cross-cutting patterns

1. I-8 lambda parameter naming is broken in most query examples. About 70 single-letter lambdas: existence_tests.py 11, populated_relations.py 16, partial_selects.py 17, aggregations.py 24, inline queries.md 12. queries.md:62 explicitly teaches "t by convention". Only predicates.py and traversal.py comply.
2. AGENTS.md I-8 is out of date with the code. col() and operator-style predicates were removed in 0.14, so the "lambda > col() > operator / slated for deprecation" rule describes nothing that exists. One page link (queries.md:874, "three predicate styles") still carries the old story. Fix AGENTS.md as well as the page.
3. Results hidden from the reader (rule 8): 27 snippets in queries.md alone. Snippet end: markers sit just above the asserts that hold the answer. The pages whose snippets include their asserts (partial_selects, aggregations) read noticeably better. Moving the markers fixes most of it.
4. No running example (rule 3) on 2 of 3 pages. models-and-fields.md uses 14 unrelated models. queries.md has 4 incompatible Transaction models plus inline references to models no schema on the page defines (Post, user.posts, user.deleted_at).
5. Schema facts drift from the DDL lowering. About 12 wrong type cells or claims in models-and-fields.md (str→TEXT, datetime→TIMESTAMP, enum→TEXT, UUID width, int→BIGINT), and singular/plural table names in queries.md (2). Generating the type table from resolve_column_storage in a test would stop this class of drift.
6. Headings mix Title Case, sentence case and slogans (rule 5): about 35 headings across the 3 pages. H2 Title Case with H3 sentence case on the same page. Sentence and slogan headings ("One verb, loud dead ends", "Zero rows, zero groups") name nothing to search for.
7. No links to the API reference (rule 4/22) on all 3 pages. api/fields.md, api/model.md and api/queries.md exist but are never linked. The pages retype option tables and limit lists instead.
8. Generated SQL is missing or simplified (rule 13). Aggregations shows no GROUP BY, models-and-fields shows DDL once out of about 10 schema features, and 4 queries.md SQL blocks are hand-simplified (unquoted reserved words, invented aliases) while the lead-ins claim "ships to the database as".

Also seen: 3 try/except-without-else example blocks that cannot fail (rule 9). Tab labels are inconsistent (1), and content.tabs.link is not enabled in zensical.toml; with it on and the labels matched, a reader's Assignment/Annotated choice would follow them across every tab set.

Cross-cutting patterns

1. mkdocstrings config renders internals and noise (9/9 API pages). show_if_no_docstring = true, show_source = true, no filters and no merge_init_into_class: about 180 docstring-less members, dunders on every dataclass, __ferro_*__ registry state, and 191 source blocks. One config change (filters, merge init, source off) plus about 60 short attribute docstrings fixes most of it.
2. Sphinx roles in Google-style docstrings (8/9 pages, about 220 literal :class:/:func:/:meth:/:data: occurrences), plus one #294 line that became a heading. Convert them to mkdocstrings cross-refs.
3. Missing public API entries (5 objects). FieldProxy, QueryNode, RelationProxy and ValueExpr are exported from ferro.query and undocumented; ModelConnection and EngineManager.session are undocumented, though they are the core of routing. Two entries use private module paths (query.builder.Query, migrations.steps.chunked).
4. Stale facts that contradict code or sibling pages (7). The architecture implementation-status note and SchemaIR bullet, lambda t: t.active (raises), the type-safety IDE completion claim, Postgres JSON instead of jsonb, the missing time row, performance's identity-map caching, and performance's "eager loading on the roadmap".
5. I-8 parameter naming (lambda t:) in predicates: 13 occurrences across architecture (3), type-safety (3), performance (3), identity-map (1) and api/queries.md table rows. The AGENTS.md I-8 text itself is stale about operator style still working.
6. Inline field-declaring examples are not test-backed (2 pages, 2 tabbed blocks). I-7 requires compile checks; tests/test_docs_examples.py covers only docs/examples/*.py.
7. Title Case and non-noun headings (6/6 concept pages, about 35 headings) against rule 5's sentence case. Several are slogans ("No Session, No Identity", "Where It Doesn't", "What This Doesn't Change").
8. Internal tracking IDs and contributor detail in user-facing text. ADR numbers (about 100 rendered mentions), #NNN, "FF-F F-2", "Phase 8.5", test-file names and a docs/plans/ link. Readers cannot act on these. Link the ADR where the reasoning helps, and drop the rest.