Add the audit, which is every quality claim in executable form

[tamnd]

Forty-one checks over the committed corpus, in six groups: structure (S01-S08), placeholders (P01-P08), glossary (G01-G06), language (L01-L08), availability (A01-A05) and hygiene (H01-H06).

Nothing here calls a model, opens a socket or reads a key. Every hard check is computable from the catalogs plus the upstream pin plus the glossary and nothing else, which is what makes it fast enough to run on every push and what makes a green run worth believing without knowing which route happened to answer that day.

Hard checks stop the build, soft ones are reported and do not. The glossary group is soft until M7. L05, L06 and L08 are soft on purpose: each has real exceptions, a heading that is genuinely an instruction or a sentence Vietnamese splits in two, and a hard rule that a correct translation breaks is a rule somebody switches off.

L04

Worth reading before the rest. The check started as "no English sentence copied through", re-deriving the classifier's verdict from the msgid and looking for English function words in whatever came back. A no-op is defined as a string with no run of two letters left once the markup is gone, so a function word cannot survive in one, and the check was incapable of producing a finding. Confirmed against the real content repo: zero hits over all 77 839 distinct msgids, by construction rather than by the corpus being clean.

Re-scoping it to literal blocks was measured and rejected: 162 candidates in the real corpus, every one of them a license text (PSF, BeOpen, zlib, W3C, WIDE, MT19937), all deliberately untranslated.

It now compares what the corpus recorded in its passthrough= provenance comment against what the classifier says today. That is the one signal that actually drifts, and only prose is reported: a string that was a no-op and is now a literal block still reads correctly, a string that is now prose is one the corpus is showing a Vietnamese reader in English.

What moved out of the library

Three things are shared rather than redefined, because a second definition is a definition that can drift from the first.

• Entry.line, the line a block started on, excluded from equality because where a string sits in a file is not part of what the string is.
• invariants.SHORT and invariants.edges, public now, because the audit asks the same two questions of a committed catalog.
• Pin.from_yaml and Pin.read, which read back what as_yaml writes and refuse anything else. S01 exists to catch a pin that has drifted from the corpus, so a reader that shrugged at a malformed file would be checking nothing at exactly the moment it mattered.

What it finds on the real corpus

Run against the content repo as it stands:

check	findings	what it is
S08	548	header drift, all files
A03	388	memory holds a machine translation the catalog never got
H04	382	file ends in more than one newline, verified with xxd
L07	306	written by gpt-5-6-mini though all three routes ask for gpt-5
S03	228	upstream drift
P07	48	doctest with a translated comment
P04	46	model added a leading space
S02	28	msgid edited here
H03	0	no key-shaped string tracked

The 306 is the one to act on and gets its own issue. All three routes request gpt-5 and 306 committed entries came back from something else.

Tests

174 new tests across seven files, one per group plus the model. Nine CLI tests for exit codes, --only, --skip, --lang, --json, --report and --fail-soft.

make check locally: 1273 passed, 97.17% coverage against an 85% floor, make secrets clean.