fix: adversarial-review runtime fixes across the telemetry + kill paths

- session-runner: exit-labeling keys off 'exit', not 'close' — an orphan
  holding the pipes could relabel a REAL exit (auth failure) as
  'timeout_startup' availability noise; the kill path still always
  group-kills and cancels the reader (labeling and unblocking are separate
  concerns). Work phase arms on a flag, not firstResponseMs===0 (a same-ms
  first byte left the startup timer live all run). The CI startup grace is
  now a real FLOOR (Math.max), matching its name and pinning test.
- gstack-detach: pgid captured AT SPAWN (== child pid under
  start_new_session) — resolving it after the grace raised ESRCH once the
  leader died on SIGTERM, orphaning TERM-immune grandchildren forever.
- test-free-shards: ledger entries carry branch + git_sha (rev-parse split:
  '--abbrev-ref HEAD HEAD' printed the branch twice and recorded it as the
  sha); local ledger default is per-PROJECT, not the machine-global tmpdir.
- eval-flake-rank: per-LINE ledger parse (one torn JSONL line vanished the
  whole series), 60-day recency bound (transcript-bearing files are MBs),
  shared isFinalizedEvalResultFile predicate (the artifact-taxonomy rule
  lived in three places); eval-store exports the predicate and finalize
  stops computing flakyRetries twice; paid-shards cleanup uses async rm
  (a SIGKILLed shard's git-workspace teardown blocked every sibling's
  stream classification on the parent event loop).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

test/helpers/eval-store.ts

@@ -192,6 +192,21 @@ export function isPartialEval(data: unknown, filename: string): boolean {
   return Boolean((data as { _partial?: unknown } | null)?._partial);
 }
 
+/**
+ * Is this path a FINALIZED eval-store result file? Single owner of the
+ * filename taxonomy (manifest.json / slice-N.json are runner artifacts,
+ * _partial* are in-progress accumulators) — the paid runner's report mode
+ * and eval-flake-rank both consume this instead of re-encoding the rule
+ * (review finding: the rule lived in three places).
+ */
+export function isFinalizedEvalResultFile(relPath: string): boolean {
+  const base = path.basename(relPath);
+  if (!base.endsWith('.json')) return false;
+  if (base === 'manifest.json' || /^slice-\d+\.json$/.test(base)) return false;
+  if (base.startsWith('_partial')) return false;
+  return true;
+}
+
 /**
  * List eval JSON files in `evalDir` plus one level of `<evalDir>/shards/<slug>/`
  * subdirectories (where the sharded paid runner points each shard's collector).
@@ -900,6 +915,7 @@ export class EvalCollector {
     const totalDuration = this.tests.reduce((s, t) => s + t.duration_ms, 0);
     const passed = this.tests.filter(t => t.passed).length;
 
+    const flaky = this.flakyRetries();
     const result: EvalResult = {
       schema_version: SCHEMA_VERSION,
       version,
@@ -917,7 +933,7 @@ export class EvalCollector {
       wall_clock_ms: Date.now() - this.createdAt,
       tests: this.tests,
       ...(this.shard ? { shard: this.shard } : {}),
-      ...(this.flakyRetries().length > 0 ? { flaky_retries: this.flakyRetries() } : {}),
+      ...(flaky.length > 0 ? { flaky_retries: flaky } : {}),
     };
 
     // Write eval file

test/helpers/session-runner.ts

@@ -162,8 +162,13 @@ export async function runSkillTest(options: {
     runId,
     env: extraEnv,
   } = options;
+  // The CI floor is a FLOOR, not a default: an explicit startupGraceMs below
+  // 300s in CI would re-open the queueing-becomes-false-red hole the floor
+  // exists for (review finding — the name promised a clamp the code lacked).
+  // Local runs honor the caller verbatim; timeout still caps everything.
+  const requestedGrace = options.startupGraceMs ?? (process.env.CI ? STARTUP_GRACE_CI_FLOOR_MS : STARTUP_GRACE_MS);
   const startupGraceMs = Math.min(
-    options.startupGraceMs ?? (process.env.CI ? STARTUP_GRACE_CI_FLOOR_MS : STARTUP_GRACE_MS),
+    process.env.CI ? Math.max(requestedGrace, STARTUP_GRACE_CI_FLOOR_MS) : requestedGrace,
     timeout,
   );
   const model = options.model ?? process.env.EVALS_MODEL ?? 'claude-sonnet-4-6';
@@ -222,9 +227,17 @@ export async function runSkillTest(options: {
   proc.stdin!.end();
   const stdoutWeb = Readable.toWeb(proc.stdout!) as ReadableStream<Uint8Array>;
   const stderrWeb = Readable.toWeb(proc.stderr!) as ReadableStream<Uint8Array>;
+  // 'exit' vs 'close' matters here: 'close' waits for stdout/stderr to
+  // drain, which an orphaned grandchild can hold open long after claude
+  // itself died with a REAL exit code — labeling must key off 'exit' or an
+  // auth failure gets triaged as 'timeout_startup' availability noise
+  // (claude adversarial finding). procExited stays 'close'-based (streams
+  // complete) for the drain race below.
+  let childExited = false;
   const procExited: Promise<number> = new Promise((resolve) => {
-    proc.on('close', (code) => resolve(code ?? 1));
-    proc.on('error', () => resolve(1));
+    proc.on('exit', () => { childExited = true; });
+    proc.on('close', (code) => { childExited = true; resolve(code ?? 1); });
+    proc.on('error', () => { childExited = true; resolve(1); });
   });
 
   // Two-phase timeout. Phase 1 (startup): no NDJSON byte yet — a shorter
@@ -240,8 +253,16 @@ export async function runSkillTest(options: {
   let phaseTimer: ReturnType<typeof setTimeout>;
 
   const killRun = (startupPhase: boolean): void => {
-    timedOut = true;
-    timedOutInStartup = startupPhase;
+    // Labeling and unblocking are SEPARATE concerns: a timer firing after
+    // the child already exited must not relabel a real exit (auth error,
+    // crash) as a timeout — but it must STILL group-kill and cancel the
+    // reader, or an orphan holding the pipes re-creates the exact
+    // blocked-drain hang this runner fixed (an early `return` here was the
+    // bug the adversarial pass caught in the first version of this guard).
+    if (!childExited) {
+      timedOut = true;
+      timedOutInStartup = startupPhase;
+    }
     // Group SIGKILL (mirrors runShardChild): claude AND every tool
     // subprocess it spawned die together — a bare proc.kill() left orphans
     // that inherited our stdout/stderr pipes and kept the API burning
@@ -264,6 +285,7 @@ export async function runSkillTest(options: {
   let liveTurnCount = 0;
   let liveToolCount = 0;
   let firstResponseMs = 0;
+  let workPhaseArmed = false;
   let lastToolTime = 0;
   let maxInterTurnMs = 0;
   const stderrPromise = new Response(stderrWeb).text();
@@ -284,7 +306,11 @@ export async function runSkillTest(options: {
         collectedLines.push(line);
 
         // Track time to first NDJSON line (measures latency from spawn to first Claude response)
-        if (firstResponseMs === 0) {
+        if (!workPhaseArmed) {
+          // Flag, not `firstResponseMs === 0`: a first line landing in the
+          // same millisecond as spawn would read as "not yet seen" and leave
+          // the startup timer live for the whole run (claude adversarial).
+          workPhaseArmed = true;
           firstResponseMs = Date.now() - startTime;
           // First byte: startup phase over — arm the work phase for the
           // REMAINING budget (total wall stays <= timeout).