Audit pipeline PR #9: determinism + fact extraction + verifier gate + KB stats + context injection (PR #9)

Bundles PR #9's work for the audit pipeline:

- N=3 consensus on cloud inference (gpt-oss:120b parallel) with qwen3-coder:480b tie-breaker
- audit_discrepancies.jsonl logs N-run disagreements
- scrum_master reviews route through llm_team fact extraction; source="scrum_review"
- Verifier-gated persistence: drops INCORRECT, keeps UNVERIFIABLE/UNCHECKED; schema_version:2
- scrum_master_reviewed flag on accepted reviews
- auditor/kb_stats.ts: on-demand observability script
- claim_parser history/proof pattern class (verified-on-PR, was-flipping, the-proven-X)
- claim_parser quoted-string guard (mirrors static.ts fix)
- fact_extractor project context injection via docs/AUDITOR_CONTEXT.md
- Fixed verifier-verdict parser to handle multiple gemma2 output formats

Empirical: 3-run determinism test on unchanged PR #9 SHA showed 7/7 warn findings stable; block count oscillation eliminated; llm_team quality scores 8-9 on context-injected extract runs.

See PR #9 for full run-by-run commit history.

auditor/claim_parser.ts

@@ -51,11 +51,20 @@ const WEAK_PATTERNS: RegExp[] = [
 
 // Empirical claims: runtime measurements / observed outcomes that can't
 // be verified from a diff (only from the actual run that produced
-// them). Example: "6/6 iterations complete, 58 cloud calls, 306s
-// end-to-end" — true, but only the test's own summary.json can
-// confirm it. Classifying as empirical lets the inference check skip
+// them). Classifying as empirical lets the inference check skip
 // diff-verification and saves the ladder for falsifiable claims.
+//
+// Two classes share this bucket because they share the skip discipline:
+//
+//   1. Runtime metrics    — "58 cloud calls", "306s end-to-end"
+//   2. History/proof refs — "verified on PR #8", "was flipping across runs"
+//
+// Both are assertions about state outside the current diff. The cloud
+// would flag them as "not backed" — but that's a false positive: the
+// proof lives in the referenced run, prior commit, or test output, not
+// in the added lines the cloud is reading.
 const EMPIRICAL_PATTERNS: RegExp[] = [
+  // ─── Runtime metrics ───
   // Iteration / attempt counts: "6/6 iterations", "attempt 5", "accepted on attempt 3"
   /\b\d+\s*\/\s*\d+\s+(iterations?|attempts?|cycles?|runs?|shards?)\b/i,
   /\b(accepted|resolved|converged)\s+on\s+attempt\s+\d+\b/i,
@@ -66,6 +75,30 @@ const EMPIRICAL_PATTERNS: RegExp[] = [
   // "escalated through N tiers", "N distinct models"
   /\bescalated\s+through\s+\d+\b/i,
   /\b\d+\s+distinct\s+(model|tier)s?\b/i,
+
+  // ─── History / proof references ───
+  // "verified on PR #8", "verified end-to-end on PR 8", "tested against PR #4"
+  // Require PR#N / commit-hash / "prior <word>" to avoid matching
+  // "verified ... in production" (PR without \b-ish anchor previously
+  // consumed "pr" of "production").
+  /\bverified\s+(?:end[- ]to[- ]end\s+)?(?:on|against|in)\s+(?:PR\s*#?\d+|commit\s+[0-9a-f]{6,}|prior\s+\w+|the\s+\w+\s+audit)\b/i,
+  /\btested\s+(?:against|in|on)\s+(?:PR\s*#?\d+|commit\s+[0-9a-f]{6,}|prior\s+\w+)\b/i,
+  // Direct PR/commit references: "PR #8", "on PR 9", "from commit abc123"
+  /\b(?:on|from|in|via|per)\s+PR\s*#?\d+\b/i,
+  /\b(?:from|in|per|against)\s+commit\s+[0-9a-f]{6,}/i,
+  // Observational descriptions of prior behavior: "was flipping", "was X before", "previously observed"
+  /\b(?:was|were)\s+(?:flipping|drifting|inconsistent|non[- ]deterministic|creeping)\b/i,
+  /\bpreviously\s+(?:observed|flagged|reported|seen|landed)\b/i,
+  /\bused\s+to\s+(?:flip|fail|flag|reject|block)\b/i,
+  /\bobserved\s+(?:in|during|on|across)\s+(?:PR|prior|\d+\s+(?:runs?|audits?))/i,
+  // "flipping/drifting across N runs" — historical variance description
+  /\b(?:flipping|drifting|varying|oscillating)\s+across\s+(?:\d+\s+)?(?:runs?|audits?|iterations?)\b/i,
+  // "the proven X" referring to prior work (proven is a STRONG pattern
+  // but in context "the proven FOO" is usually a historical reference,
+  // not a fresh claim). We catch it here so the empirical skip wins.
+  /\bthe\s+proven\s+(?:escalation\s+ladder|pipeline|flow|loop|tier|path)/i,
+  // "from the 9-run test", "across the 5-run validation"
+  /\b(?:from|across|in|during)\s+the\s+\d+[- ]run\s+(?:test|validation|probe|experiment)/i,
 ];
 
 export interface ParsedClaims {
@@ -101,7 +134,7 @@ function scanText(text: string, location_prefix: string, commit_sha: string, out
     // classify it as empirical so the inference check doesn't ask
     // the cloud to prove "58 cloud calls" from the diff. Order:
     // empirical → strong → moderate → weak.
-    const empirical = firstMatch(line, EMPIRICAL_PATTERNS);
+    const empirical = firstUnquotedMatch(line, EMPIRICAL_PATTERNS);
     if (empirical) {
       out.push({
         text: line.trim().slice(0, 200),
@@ -111,7 +144,7 @@ function scanText(text: string, location_prefix: string, commit_sha: string, out
       });
       continue;
     }
-    const strong = firstMatch(line, STRONG_PATTERNS);
+    const strong = firstUnquotedMatch(line, STRONG_PATTERNS);
     if (strong) {
       out.push({
         text: line.trim().slice(0, 200),
@@ -121,7 +154,7 @@ function scanText(text: string, location_prefix: string, commit_sha: string, out
       });
       continue;
     }
-    const moderate = firstMatch(line, MODERATE_PATTERNS);
+    const moderate = firstUnquotedMatch(line, MODERATE_PATTERNS);
     if (moderate) {
       out.push({
         text: line.trim().slice(0, 200),
@@ -131,7 +164,7 @@ function scanText(text: string, location_prefix: string, commit_sha: string, out
       });
       continue;
     }
-    const weak = firstMatch(line, WEAK_PATTERNS);
+    const weak = firstUnquotedMatch(line, WEAK_PATTERNS);
     if (weak) {
       out.push({
         text: line.trim().slice(0, 200),
@@ -143,9 +176,35 @@ function scanText(text: string, location_prefix: string, commit_sha: string, out
   }
 }
 
-function firstMatch(text: string, patterns: RegExp[]): RegExp | null {
+// Match a pattern only when its match position is NOT inside a quoted
+// string on the line. Mirrors the same guard in auditor/checks/static.ts
+// — the two files have the same false-positive class: PR authors
+// quote pattern examples in commit message bodies (e.g. `"Phase 45
+// shipped"` as a test example) and without this guard those quoted
+// references get flagged as fresh ship-claims. Only skips when the
+// match itself falls inside quotes; real (unquoted) uses of the same
+// vocabulary still classify correctly.
+function firstUnquotedMatch(text: string, patterns: RegExp[]): RegExp | null {
   for (const p of patterns) {
-    if (p.test(text)) return p;
+    const m = text.match(p);
+    if (!m || typeof m.index !== "number") continue;
+    if (isInsideQuotedString(text, m.index)) continue;
+    return p;
   }
   return null;
 }
+
+// Walks left→right toggling in-quote state on each unescaped quote.
+// Good enough for single-line claims; multi-line strings aren't parsed.
+function isInsideQuotedString(line: string, pos: number): boolean {
+  let inDouble = false, inSingle = false, inBacktick = false;
+  for (let i = 0; i < pos; i++) {
+    const c = line[i];
+    const esc = i > 0 && line[i - 1] === "\\";
+    if (esc) continue;
+    if (c === '"' && !inSingle && !inBacktick) inDouble = !inDouble;
+    else if (c === "'" && !inDouble && !inBacktick) inSingle = !inSingle;
+    else if (c === "`" && !inDouble && !inSingle) inBacktick = !inBacktick;
+  }
+  return inDouble || inSingle || inBacktick;
+}