feat(spine): Gate-1 wind-tunnel — wire diff-replay into un + S4 hard-error re-derivation + verdict-precedence ratification (post-#2188 deferrals)

[satur8d]

Context

#2188 shipped the Gate-1 wind-tunnel harness (windtunnel.lock.v1 schema + pure scorer + git-derived freeze/integrity proofs + the post-image readStrategy parity seam proven at the unit level + mock-engine validation), per the OQ2-ratified harness-only-now scope. Three items were deliberately deferred and are tracked here so they are not lost. They land with the post-strategy#516 certifying-run build.

Deferred items

1. Wire the shared post-image readStrategy into the run command's engine path (C1/S1 integration).
   Today run builds the readStrategy and hands it to a mock engine (harness phase — no compiled rules yet). The certifying run must build additions from each resolved PR's diff and feed them through enrichWithAstContext + applyAstRulesToAdditions with the one shared post-image readStrategy (the mechanism is already unit-proven in packages/core/src/spine/windtunnel-parity.test.ts). Blocked on strategy#516 (real minted rules to replay).

2. S4 full hard-error corpus-completeness re-derivation.
   Today freeze verifies lc-clone accessibility + asOfCommit ancestry + counts/warns on resolvedPrs. The spec's S4 contract is a hard-error assertion resolvedPrs === selectionRule(asOfCommit) re-derived from lc merge history. Implement the deterministic selectionRule(asOfCommit) resolver (code-touching predicate per ADR-110 — conventional-commit type ∈ {feat,fix,refactor,perf} OR source-glob, excluding chore(deps)/docs) and make any mismatch a hard error (no silent shrink, §5).

3. Ratify the scorer verdict precedence.
   scoreWindtunnel currently ranks exposure-floor + cull-rate HONEST-NEGATIVE above FP → FAIL. The spec's "any firing labeled FP ⇒ FAIL" reads as unconditional, so a confirmed FP co-occurring with thin exposure is currently masked as HONEST-NEGATIVE. Ratify the intended precedence (likely FP→FAIL wins) and lock it with a test. Strategy-claude / operator adjudication.

Refs

#2188 · ADR-110 (mmnto-ai/totem-strategy#669) · Proposal 299 (strategy#661) · strategy#516

[satur8d]

Reopened — PR #2193 shipped item 3 only (precision sentinel + verdict precedence). This issue was auto-closed by the merge because the PR body literally contained the phrase \close #2189\ (in a sentence stating it does NOT close it) — GitHub's keyword parser ignores the negation. Items 1 (wire readStrategy into
un\ — blocked on strategy#516) and 2 (S4 selectionRule corpus re-derivation) remain open.