Parse joint and concurrent resolutions, and keep their preamble recitals

[Nitjsefnie]

Related issue

Closes #201

What does this change?

Joint and concurrent resolutions parse instead of raising, and every resolution form is labelled
correctly.

Three parts, in the order they matter:

1. The body. find_bill_body now recognises <resolution-body>, which is what made
   normalize_bill raise on every resolution.
2. The label. _LEGIS_NUM_RE captures the whole prefix and _extract_metadata normalises it, so
   all eight forms map correctly. Before this, j and n were chamber-ambiguous rather than
   merely wrong — hjres and sjres were indistinguishable, as were hconres and sconres.
3. The preamble. Recitals are captured as one front-matter node keyed "preamble", ordered before
   the enacting clause, riding the extract_front_matter_nodes mechanism XML diff: bill front matter (form block + enacting clause) missing from full-bill text and diff #48 added for exactly this
   ("content outside the body was dropped from the full-bill text and the diff").

Your trap-at-pickup note was right, and it went one layer deeper than the issue. Fixing the body
finder alone drops the recitals of the ~14% of resolutions carrying a preamble — that is handled
above. But reported-stage resolutions carry something else: paired committee-amendment variants,
two <resolution-body> and sometimes two <preamble> children, struck and amended. root.find()
takes the first, which is the struck one, so the fix as first written parsed those clean and rendered
the superseded text as the document — turning your loud ValueError into a quiet wrong answer.
They now raise, naming what was found:

ValueError: Resolution carries paired committee-amendment variants (2 <resolution-body>,
2 <preamble>: resolution-body [deleted, added]; preamble [deleted, added]). Choosing between
the struck and the amended text is not supported.

Choosing between the variants is a real feature and your call, so this refuses rather than guesses.
Nothing regresses: those files already fail today.

Two things the sweep turned up that shaped the guard. The shape varies — of the three affected
files, one has 2 bodies and 0 preambles, one 2 and 1, one 2 and 2 — so a preamble-only guard would
miss two of three, and both child kinds are counted independently. And the struck block is often
unmarked: in two of three the first <resolution-body> carries no changed attribute at all, so
a future variant-selection feature cannot key on that attribute either.

The issue's mapping table is incomplete, worth knowing since it will be the reference next time:
simple resolutions (H. RES. / S. RES.) share the <resolution> root, so the body fix makes them
parse and then stamps them with the bill designators H.R. / S. on a real document. All eight
forms are handled and tested.

How to test

Prevalence was measured, not estimated: all six 119/1 type archives were downloaded from govinfo and
all 2,122 files parsed. Exactly 3 of the 62 reported-stage files carry paired variants
(hres137rh, hres237rh, sres226rs), and those same 3 are the only ones flagged across every
stage.

Fixtures are four real GPO XML files, committed, each byte-identical to what govinfo serves today.

Bite, in both directions:

• The preamble tests are proven against the body-fix-only state, not against stock develop —
  stock fails them for the wrong reason (the crash) and would prove nothing. With only the 5-line body
  fix applied, hconres 58 parses to exactly 4 nodes with 0 of its 12 recitals; the tests fail on
  absence.
• Gutting the variant guard gives DID NOT RAISE ValueError — the silent-wrong-answer state.
• Gutting the normalization to a bare .lower() turns 4 red, e.g.
  'H. CON. RES. 58' -> 'h. con. res.' is not letters-only.

The 30 distinct <legis-num> spellings in the committed corpus are unchanged by the regex swap,
including the unspaced H.R. 2029 form; the only three that change are the new resolution fixtures.

Gates: uv run ruff check . clean · uv run ruff format --check . 178 files · fast
1417 passed, 15 skipped, 15 xfailed against a 1403/15/15 baseline · browser 6 passed · slow
564 passed, 7 skipped against 554/7. Skips unchanged in both tiers, so nothing needed an allowlist
entry. Those counts are from this machine and run higher than yours — offered as a delta.

A warning-count delta is deliberately not quoted: three runs of identical code gave 4, 5 and 3, since
xdist dedupes per worker.

Checklist

• Linked the issue above (Closes #...)
• Ran the CI gates locally and they pass (see What CI checks)
• New or changed behavior has tests
• For a bug fix: the test fails without the fix, and I ran it both ways to check
• Disclosed AI assistance below, if any

AI assistance

Generated by Claude Opus 5 (1M context) (brief, implementation, review), Kimi K3 (verification)

Follow-ups / Known Limitations

• The resolving-clause label is now more visible than it was. Resolutions used to crash, so nobody
  saw that they get the bill's "Be it enacted…" boilerplate. Out of scope here as the issue scopes it,
  but it is the natural next change and I am happy to take it.
• No second guard was added in the front-matter path, deliberately: normalize_bill calls the body
  finder first and grep confirms those are the only callers, so a second guard would be unreachable —
  which is the same thing you had us remove on fix(fetch): verify the archive opens before caching a download #211.
• The preamble is only reachable when a <form> element exists, inherited from XML diff: bill front matter (form block + enacting clause) missing from full-bill text and diff #48's early return. All
  149 real 119/1 resolutions sampled have one, and none has a preamble without a form, so the gap is
  bounded — but it is XML diff: bill front matter (form block + enacting clause) missing from full-bill text and diff #48's contract, not something to change here.
• Per-recital preamble nodes, the HTTP 422 message in server/app.py, and the PDF path are all out of
  scope and yours to call.