docs(review-rigor): V3 — prevention does not close a detection finding

[ClintonSarkar]

Closes #271

Adds V3 to core/review-rigor → Verdict integrity: a finding about detection is not closed
by a change to prevention.

Why V1 doesn't already cover it

V1 bans speculative clears — it keys on hedge-words and unverified claims. The miss this comes from
had neither. A self-review pass correctly found "nothing gates this invariant against regrowth", the
author answered it with a declarative constraint that pins the invariant, and that constraint was
real, working, better code. Nothing was hedged; the item was marked resolved with confidence.

But the finding was about verification, and the answer was about likelihood. Declared is not
checked — deleting the declaration broke nothing, so the gap the finding named survived. Two further
review passes then inherited the disposition rather than re-litigating it, which is the documented
behaviour of Re-review continuity, and the human reviewer's first response was
CHANGES_REQUESTED on exactly that gap.

Self-review is especially exposed here, because the accepted resolution is frequently the reviewer's
own earlier suggestion — there is no independent party to notice the goalposts moved.

The rule as encoded

When the finding is that something is unverified — no test, no assertion, nothing fails if it
regresses — it closes ONLY when something FAILS on regression.

The test that settles it: if someone deletes this hardening, what turns red? "Nothing" means the
finding stands.

with the detection cue (the finding's own words are no test / nothing catches / unenforced / could
silently regress, and the resolution changed a mechanism instead of adding an observer), the
self-review/re-review exposure, and a ✅/❌ pair.

It also carries one sub-rule learned expensively in the same incident: a red-then-green demo is only
evidence if the artifact under test actually moved. Two negatives passed vacuously before a valid
one existed — the first widened a declarative allowlist whose symbols a different flag already
suppressed, so the built output was byte-identical; the second used a pattern that silently failed to
match because of line-ending differences, so the change was never applied and the build stayed green.
Both read exactly like a passing check. Editing the source that supposedly produces an artifact is not
the same as changing the artifact.

Scope

docs/core/review-rigor.md only, +10/−4:

• V3 added after V2 in Verdict integrity; the section heading and the changes-verdict line updated
  from V1–V2 to V1–V3.
• Re-review continuity — the verdict-integrity sentence now names V3 and says a resolution accepted
  in an earlier pass is inherited rather than re-litigated, so re-ask closed-vs-mitigated.
• Tests — the "is there a regression test?" bullet now says hardening is not a substitute, pointing
  at V3.

Swept for stale references to the old V1–V2 range (none remain) and for Go-embedded doctrine
referencing these rules (none), so this is a docs-only writeback with no toolchain verification
required.

---

🤖 Generated with Claude Code

📚 Learned by detritus — absorb session/2026-08-06-export-surface-invariant