Let the card check tell a site that grew from a wrong card

[realmarcin]

Closes #148, closes #113, closes #115, closes #176, closes #114, closes #217, closes #218, closes #219, closes #220, closes #221, closes #222, closes #231, closes #232, closes #233, closes #234, closes #235, closes #239, closes #240, closes #241, closes #242, closes #243, closes #247, closes #248, closes #249, closes #250, closes #257, closes #258, closes #259, closes #260, closes #261, closes #262, closes #263, closes #264.

What changes

scripts/fleet/check_cards.py sorts each card into one verdict, judged against _fleet/data/site_audit.json, which the refresh writes:

Verdict	When	Exit
ok	site equals card	pass
grew	site ahead of card, at most 14 days after the audit's pinned_at_utc (GRACE_DAYS), and at most 50% ahead (MAX_LEAD)	warning
STALE	site ahead past either limit, or no audit	fail
WRONG	card differs from figure_at_pin, the figure the audit read from the source's committed copy at the pin	fail (#217, #231)
SHRANK	site behind card	fail
GONE	source answered a 4xx other than a throttle	fail (#176)
CHANGED	source arrived with no readable figure, or CultureMech's generated block is missing	fail (#176)
MARKUP	a card without exactly one readable headline tile, a figure outside every card, or two cards for one Mech	fail (#218, #261, #262)
UNCARDED	a card with no SOURCES entry, or the reverse	fail
AUDIT	site_audit.json missing or malformed (including null, {} or no repositories), a figure_at_pin missing or not a whole number, or pinned_at_utc missing, unreadable or in the future (a time with no offset is taken as UTC)	fail (#232, #242, #250)
unread	DNS, timeout, dropped connection, 5xx, or a 408/425/429 throttle	warning (#219, #220)
UNCHECKED	more than half the sources unread	fail (#115)

• check_cards.py cannot tell a site that grew past a pinned card from a wrong card #148, The nightly card check stays red for Mechs that publish after a refresh's pins #113, The 10% growth tolerance goes red a day after a refresh, and passes a card mistyped inside 10% #217: the page is a snapshot at a refresh's pins, and fast Mechs publish within hours of them. A fixed 10% allowance (round 1 of this PR) went red a day after every refresh, because CellStructureMech adds about 9% a day. It also passed a card mistyped inside 10%.

  • Growth now warns for a fortnight, and fails sooner once a site is half as large again as its card.
  • A wrong card fails against figure_at_pin, which site_audit.json now records for the nine sources with a committed copy. A test requires it to equal each card, so a typo fails on the PR, and the nightly reports it however far the site has moved. Round 2's version compared the source's hash with its pinned copy, which stopped working as soon as a source changed at all (check_cards: WRONG fires only while a source is byte-identical to its pin, so a mistyped card on a moved source passes #231).

• check_cards.py passes when a source returns 404 or its wording changes, against its own docstring #176: CultureMech's figure is read only between its GENERATED CORPUS STATS markers (REGIONS).

• The card markup is parsed by three independent regexes #114, check_cards: an extra card figure is accepted, and a card with an unreadable figure is reported as missing #218: the new scripts/fleet/card_markup.py is the one card parser. assemble_page.py refuses a card without exactly one figure, as it did before this PR. A card with an unreadable figure is reported as MARKUP, not as a missing card.

• Docs: _fleet/README.md and both skills are updated. Step 7 of the update skill now derives audit figures through read_source(), which applies REGIONS (update skill step 7 derives the audit figure through check_cards.published(), which skips REGIONS #221). Step 8 says what grew and STALE mean.

• The nightly now runs the card check even when an earlier step failed. A mistyped card fails the provenance tests first, and the card report, with its WRONG line and every other card's verdict, still prints (The nightly skips check_cards.py whenever a unit test fails, so WRONG and AUDIT never print in CI #239). WRONG's remedy is to correct every occurrence of the figure, found by grepping the tree for it, not a refresh (check_cards: the closing line sends a WRONG card to a full refresh; the README gives it its own remedy #240, check_cards: WRONG's remedy names only the card, but its figure is repeated in the MECHS block and on other pages #248, check_cards: WRONG's remedy claims to name every copy of a card figure but misses MECHS extra: prose and cross-reference text #258).

• Review history: five internal review rounds, then a Codex review, which found check_cards: a figure_at_pin that is a string or float, or missing, silently switches WRONG off #260 to review-open-issues skill says a growing card only warns for 14 days, leaving out the 50% lead limit #263, then a verification round, which found card_markup accepts a figure with misplaced commas, such as 32,06 #264. All are fixed here.

Evidence

• python3 -m unittest discover -s tests: 74 tests pass. That includes a new test holding each MECHS records: to its card (check_cards: WRONG's remedy names only the card, but its figure is repeated in the MECHS block and on other pages #248). CardCheckTests covers every verdict, the grace and lead boundaries, 429/408/425/503, a truncated body, a json source serving HTML, the 5-of-10 and 6-of-10 unread boundary, a half-marked region, three markup defects, and main()'s exit code (check_cards tests miss the parse-exception path, the more-than-half boundary and a half-marked region #222).
• python3 scripts/fleet/assemble_page.py --check: mechs.md is unchanged.
• Live run, 2026-09-25, 18 hours after the pins:

    grew      CellStructureMech    card 542, site 580 (+38 in the 18 hours since the pins)
    grew      TraitMech            card 763, site 768 (+5 in the 18 hours since the pins)
  2 grew, 8 ok.
  Passing with warnings; see above.
  

Judgement calls

• The grace period and lead limit (14 days, 50%). While any Mech grows, the nightly asks for a full refresh a fortnight after the pins, or sooner for a Mech as fast as CellStructureMech (under a week). Both are constants at the top of the script.
• ProteinTraitsMech has no figure_at_pin. Its data file is built in CI, so there is no committed copy at the pin. Its card is checked only against the live site.
• figure_at_pin in the current audit was read from the committed copies at the Refresh the X-Mech page from the Mechs as they are now, and add the skill that does it #120 pins in the refresh snapshot. Each copy's sha256 matched the audit's recorded *_sha256_at_pin, and each figure equals its card.

🤖 Generated with Claude Code