feat: a gate that asks whether the documentation tells the truth about the tree (audit claim F-06/F-07)

[sinkog]

Nothing asked that until an external audit did — and what it found was worse than stale numbers.

document	said	tree
SPEC.md (the declared authority)	"normative, one implementation"	two
SPEC.md	"executed by one implementation"	both
conformance/README.md	"No implementation exists … Nothing here has passed"	both run the whole corpus
docs/spec-vector-map.md	Rust "does not exist yet"; agreement "still an untested claim"	byte-identical on every vector
Makefile	"mk/rust.mk is NOT present"	included four lines below
README.md	44 invariants	46
docs/en/architecture.md	34 invariants, 27 vectors	46, 37

None of that was carelessness in the ordinary sense. Every one of those sentences was true when written, and nothing connected it to what it described.

A count copied into prose is a claim with no owner: it does not fail when it stops being true, it just stops being true.

The gate

Measures the tree — invariants from the §12 register, vectors from the corpus, implementations from has sources and has a conformance runner (a directory with a .gitkeep is not an implementation) — and compares it to what documents say.

Two kinds of check, because there are two ways a status claim goes wrong:

• numbers — derived and compared
• situations — "one implementation", "never executed" — which cannot be derived, so they are listed and refused when the measurement contradicts them

Verified it can fail: 46 → 44 in README fails; two implementations → one implementation in SPEC fails; adding one vector directory fails two claims at once.

Two limits, printed in the tool's own output

It checks the claims listed in it and nothing else. Prose in a file it does not know about drifts silently — exactly as all of the above did. The lists are explicit rather than clever for that reason: every entry is a claim someone decided was worth binding.

The phrases are deliberately long. "one implementation" also matched "without the corpus and the other implementation noticing" and a sentence about why 0.1's bootstrap rule deadlocked — both correct prose. A check that fires on legitimate text gets switched off, and then it checks nothing.

docs/hu/architecture.md is registered for no count at all, because it states none. A document that makes no numeric claim has nothing to go stale — the cheapest fix available for this whole class.

---

make ci passes; status-claims runs in ci.gates. 10 of 37 audit findings closed.