valent-pipeline 0.19.33 → 0.19.34

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "valent-pipeline",
-  "version": "0.19.33",
+  "version": "0.19.34",
   "description": "v3 multi-agent AI pipeline for software development lifecycle",
   "type": "module",
   "bin": {

package/pipeline/orchestrators/claude-code/sprint.workflow.js

@@ -2833,8 +2833,9 @@ async function runStory(story) {
     const evidenceNote = hasMachineEvidence
       ? `\n\nMACHINE EVIDENCE (authoritative): the deterministic Evidence gate already ran ` +
         `\`evidence assert\` + \`trace check\` (gate verdict: ${evGate.verdict}${evGate.escalated ? ', escalated after the rework cap' : ''}${evGate.advisory ? ', advisory' : ''}). ` +
-        `Read the NEWEST JSON under \`stories/${sid}/evidence/asserts/\`` +
-        `${evGate.verdictPath ? ` (the gate reported \`${evGate.verdictPath}\`)` : ''} and TRANSCRIBE testsPassed/testsFailed/openP1toP3 ` +
+        (evGate.verdictPath
+          ? `Read the assert JSON the Evidence gate VERIFIED — \`${evGate.verdictPath}\` — NOT merely the newest file under \`stories/${sid}/evidence/asserts/\` (a transient or mislabeled-label run can be newer and carry the wrong counts) — and TRANSCRIBE testsPassed/testsFailed/openP1toP3 `
+          : `Read the NEWEST green qa-b JSON under \`stories/${sid}/evidence/asserts/\` and TRANSCRIBE testsPassed/testsFailed/openP1toP3 `) +
         `from its counts (openP1toP3 = bugs in the \`valent:bugs\` block of bugs.md with status open/fix-in-progress AND priority P1/P2/P3, EXCLUDING any with \`preExisting: true\` or \`category: pre-existing\` — baselined pre-existing bugs are non-blocking and the cross-check counts them separately, so including them here produces a verdict-integrity mismatch and refuses the ship) — do NOT count from execution-report.md prose. ` +
         `Set evidenceVerdict to the assert verdict and gitSha to the evidence run's SHA. Your job on numbers is ANOMALY ` +
         `DETECTION: FLAG as a discrepancy any claim in execution-report.md / traceability-matrix.md / bugs.md that the ` +

package/src/lib/crosscheck.js

@@ -62,17 +62,31 @@ export function crosscheckVerdict({ reported = null, bugsBlock = null, assertVer
   // ship is claimed < actual — and refusing it stranded genuinely shippable stories. So an
   // over-report on these fields reconciles to the authoritative recount with a warning; an
   // under-report stays a hard mismatch. (Prompt tightening alone did not hold — the LLM still
-  // over-counts; this is the mechanical backstop.) Good-thing counts (testsPassed) and exact-match
-  // fields keep strict equality, where over-claiming IS the unsafe direction.
+  // over-counts; this is the mechanical backstop.)
+  //
+  // GOOD-thing counts (testsPassed) are the MIRROR IMAGE: the unsafe, ship-sneaking direction is
+  // OVER-claiming (padding green to look more passing than reality), so an over-claim stays a hard
+  // mismatch; an UNDER-report (claimed FEWER passes than the recount) is the safe direction — you
+  // cannot sneak a ship by under-stating passes, because the ship gates on the authoritative recount,
+  // not on this transcribed number — so it reconciles to the recount with a warning. This closes the
+  // JUDGE-verdict-path analogue of Gap 2 (label-multiplexing, 2026-06-18): a transcribed testsPassed
+  // bound to a transient/mislabeled qa-b run under-counts passes, which is harmless and must not
+  // refuse a ship the green assert already proved. Exact-match fields (verdict, reviewedSha) keep
+  // strict equality in both directions.
   const SAFE_IF_OVER_REPORTED = new Set(['testsFailed', 'openP1toP3']);
+  const SAFE_IF_UNDER_REPORTED = new Set(['testsPassed']);
   const compare = (field, actual) => {
     recounted[field] = actual;
     if (!claim(field) || reported[field] === actual) return;
-    if (SAFE_IF_OVER_REPORTED.has(field) && typeof reported[field] === 'number'
-        && typeof actual === 'number' && reported[field] > actual) {
+    const numeric = typeof reported[field] === 'number' && typeof actual === 'number';
+    if (numeric && SAFE_IF_OVER_REPORTED.has(field) && reported[field] > actual) {
       warnings.push(`${field}: self-reported ${reported[field]} but artifacts show ${actual} — over-report (safe, conservative direction), reconciled to the recount`);
       return;
     }
+    if (numeric && SAFE_IF_UNDER_REPORTED.has(field) && reported[field] < actual) {
+      warnings.push(`${field}: self-reported ${reported[field]} but artifacts show ${actual} — under-report (safe direction: under-stating passes cannot sneak a ship), reconciled to the recount`);
+      return;
+    }
     mismatches.push({ field, claimed: reported[field], actual });
   };