Check the measure ids the US legacy ETL cleanup doc names

[MaxGhenis]

Summary

test_us_legacy_etl_cleanup_doc_accounts_for_retired_series (added in #348) checked only that the backticked package ids in docs/us-legacy-etl-cleanup.md are shipped packages. It did not check the measure ids written next to them, so a renamed or mistyped measure id passed CI. An independent review of #350, which edits five of those rows, flagged the gap. Two ids had already drifted on main:

• The mortgage row named soi-congressional-district-2022 interest_paid_deduction_*. chronicle#345 renamed those measures to mortgage_interest_paid_* (still N19300/A19300; see docs/concept-migrations.md, "Schedule A line 8a relabel").
• The SSI row put ssi_payments after three packages. It is a row value id of ssa-annual-statistical-supplement-2025's payment_amount measure. ssa-ssi-table-7b1-2024 and ssa-ssi-monthly-statistics-2024-12 declare neither.

Test. The test now parses every "Chronicle package and measures" cell under a grammar stated in a comment above the helpers:

• A cell splits into clauses at each ; outside parentheses.
• Within a clause, a backticked id belongs to every package listed since the previous id. So `a`, `b`: `x` claims x for both packages, and `a` `x`, `y` claims both ids for a.
• An exact id must be a measure id or a record-set row value id that the package declares. A prefix_* wildcard must match at least one of its measure ids.
• An IRS column code (N07225, A06500, ...) must be a source_column_id read by a measure that the id before it matches.
• A token that is none of these kinds fails the test, and so does an id with no package before it.

How tables are read. Three adversarial review rounds each found a valid Markdown edit that put a bad claim on the rendered page while the test passed:

• round 1: an indented row, a row without outer pipes, an escaped \|;
• round 2: a second table under a bold header, a column named twice;
• round 3: a table inside a blockquote, list item or <details> block, and an extra table headed Chronicle package and measures\*.

The test now reads the doc with markdown-it-py, which is CommonMark plus GitHub's table extension. A table therefore counts wherever GitHub renders one, under a heading of any form. Every column whose header reads "Chronicle package and measures" once emphasis, code and escape marks are stripped is checked. DOC_TABLES pins every table's heading and header, so an added, moved or reheaded table fails until it is listed. Every reproduction from the three rounds is now a real-doc mutation test. markdown-it-py was already locked at 4.0.0 through rich. This PR declares it as a dev dependency, and the uv.lock diff is metadata only.

The test reads each package's YAML through chronicle.source_package.load_source_package and builds no facts. On main's doc it reports 5 errors: the renamed district measure and its two columns, plus ssi_payments against the two SSI packages that lack it. With this PR's doc it checks 64 claims (52 ids, 12 column codes) and finds none. Packages listed together are each checked: all three ACA OEP packages, both Medicaid packages and the three packages that publish net_capital_gains_*.

New tests:

• test_us_legacy_etl_cleanup_doc_check_flags_edited_claims applies 10 edits to the real doc. One renames a measure; the rest are the review reproductions. Each must raise exactly the named errors.
• test_us_legacy_etl_cleanup_doc_check_reads_any_heading_form checks that indented, closed and setext headings raise no error.
• test_markdown_tables_read_cells_as_github_renders_them and test_replacement_cells_read_every_replacement_column cover the reader.
• test_replacement_claim_check_flags_undeclared_ids runs one-row fake tables against fake packages. It covers a nonexistent measure id, a wildcard with no match, a wildcard that would match only a row, a listed package that lacks the id, a wrong column code, a column code after a row id, an unknown package, an id or column with nothing before it, an id cut off by ;, and an unreadable token.
• test_replacement_claim_check_accepts_declared_ids covers the forms the doc uses. It pins the claims each one produces.

Doc. This PR fixes the two rows. It also rewrites the QBI gap note, which still said the broad package's qbi_* read N03270/A03270 "on main". chronicle#345 repinned them, and main's package reads N04475/A04475. The new wording is #350's, character for character, so that line merges cleanly.

Interaction with #350. A trial merge of this branch onto #350's head (a838ee45) conflicts in one hunk: the SALT, mortgage and charitable rows, which #350 rewrites. Taking #350's side resolves it. In the resolved tree, #350's new rows (taxes_paid_deduction_*, broad mortgage_interest_paid_*, charitable_* and taxable_state_local_income_tax_refunds_*, with their column codes) pass the new check, and all 19 tests in this group pass.

Chronicle Governance

This PR does not touch the Chronicle source-data boundary. It adds no source package, target profile, consumer contract or schema, and it changes no fact. It changes a boundary test, a doc, a changelog fragment and the dev dependency list.

• Approved Chronicle agent role: n/a (no source package, target profile or contract change)
• Deterministic checks run: uv run ruff check tests/test_chronicle_boundaries.py; uv run pytest tests/test_chronicle_boundaries.py -k "cleanup_doc or replacement_ or markdown_tables" (33 passed)
• LLM judge verdicts: n/a (no source-data boundary change)

Tests

• uv run pytest -q tests/test_chronicle_boundaries.py -k "cleanup_doc or replacement_ or markdown_tables": 33 passed, about 5s locally. The doc test takes about 4s because it parses the named packages' YAML. The other tests reuse that parse through a cache.
• Same selection on main's doc: the tightened test fails with the 5 errors above.
• Same selection on a trial merge with Add IRS Historic Table 2 state facts for refunds, taxes paid, mortgage interest and charitable contributions #350 (conflict resolved to Add IRS Historic Table 2 state facts for refunds, taxes paid, mortgage interest and charitable contributions #350's rows): 19 passed at the first head. Since then I have checked that the reader yields the same 30 cells as the original reader on main's and on Add IRS Historic Table 2 state facts for refunds, taxes paid, mortgage interest and charitable contributions #350's docs, and that Add IRS Historic Table 2 state facts for refunds, taxes paid, mortgage interest and charitable contributions #350's table layout matches DOC_TABLES.

🤖 Generated with Claude Code

[MaxGhenis]

Independent review trail. Each round was GPT-6.1 Sol, adversarial, run through Subfleet.

Round	Head	Verdict	Blocking findings	Fixed in
1	c0f750f	REQUEST_CHANGES	Valid Markdown skipped or split rows: an indented row, a row without outer pipes, an escaped |	44f43f3
2	44f43f3	REQUEST_CHANGES	A second table under a bold header was skipped, and a repeated column was half-checked	309dc8a
3	309dc8a	REQUEST_CHANGES	A table inside a blockquote, list item or <details> block was skipped	ebffab4 (markdown-it-py reader, DOC_TABLES pin)
4	ebffab4	APPROVE	none	—

Every reproduction from rounds 1-3 is now a case in DOC_MUTATIONS.

Round 4 left one nit: a prose line placed directly under a table, with no blank line between, fails the doc test. That is the correct result, not a false failure. GitHub's own GFM spec (cmark-gfm test/spec.txt, the example after "The table is broken at the first empty line") absorbs such a line as an extra table row. GitHub would therefore render it as a row with an empty "Chronicle package and measures" cell, and markdown-it-py reads it the same way. The cmark-gfm extensions.txt fixture cited in that round has a blank line before Hi!, which is why it ends the table there.

Round 4's sandbox could not run the lock check. I ran uv lock --check locally and it passes; the uv.lock diff is metadata only.

Merge note for chronicle#350: whichever PR merges second has one conflicting hunk in docs/us-legacy-etl-cleanup.md, the SALT, mortgage and charitable rows. Take #350's version of those rows. The reader returns the same 30 cells and the same table layout on #350's doc, and #350's new rows pass the check.