claude-code-session-manager 0.98.0 → 0.101.0

package/src/main/__tests__/runVerify-landed-commit-outranks.test.cjs

@@ -1,191 +0,0 @@
-/**
- * runVerify-landed-commit-outranks.test.cjs — ground truth (exit 0 + a commit
- * landed THIS run + no explicit FAIL sentinel) outranks the transcript-
- * pattern heuristics (final-20% is_error, FAIL/Traceback,
- * ModuleNotFoundError), so a self-recovered tool error no longer parks a
- * green, committed run.
- *
- * Fixture: the REAL run log for 1218-fo-01-move-scripts-lib-into-src-main-lib
- * (session-manager-operations/reviews/2026-09-13-scheduler-stability-
- * investigation.md), copied verbatim from
- * ~/.claude/session-manager/scheduled-plans/runs/2026-09-13T21-52-51-192Z/ —
- * a `git apply` "patch does not apply" tool error at event 179/209 (inside
- * the final 20% of the transcript) that the agent retried and resolved 3
- * events later, in a run that landed commit d1edf15 and was STILL parked
- * needs_review (verdicts.json: verdict transcript_errors, reason
- * "is_error=true in final 20% of transcript (event 179/209)"), spawning a
- * $3.74 fix-plan investigation into work that needed no fix.
- *
- * The real run's own SCHEDULER_VERDICT sentinel is an explicit FAIL — an
- * honest, UNRELATED report that test:unit was red for proven pre-existing/
- * environmental reasons (shared /tmp contention), not a regression from this
- * PRD's move. Per this PRD's own AC, an explicit FAIL sentinel must keep
- * today's behavior byte-for-byte, so scenarios (a)/(b)/(d) below use a
- * variant of the fixture with that trailing SCHEDULER_VERDICT line stripped
- * (sentinel === null) — the majority real-world park shape this rule targets
- * (the findings report measured 77 of 212 parks as transcript_errors with NO
- * sentinel at all, vs. this one specific row's own additional, unrelated
- * FAIL claim). Scenario (c) reuses the pristine, byte-for-byte-unmodified
- * fixture to prove the real, still-present FAIL sentinel is honored exactly
- * as before this change.
- *
- * Run standalone: timeout 120 npx vitest run src/main/__tests__/runVerify-landed-commit-outranks.test.cjs
- */
-
-import { test, beforeAll, afterAll, vi } from 'vitest';
-const assert = require('node:assert/strict');
-const os = require('node:os');
-const fs = require('node:fs');
-const path = require('node:path');
-const { execFileSync } = require('node:child_process');
-const { verifyRun } = require('../runVerify.cjs');
-
-// Sidecar goes through config.writeJsonSync, whose allow-list rejects os.tmpdir() fixtures.
-let writeSpy;
-beforeAll(() => {
-  writeSpy = vi.spyOn(require('../config.cjs'), 'writeJsonSync').mockImplementation((abs, data) => {
-    require('node:fs').writeFileSync(abs, JSON.stringify(data, null, 2) + '\n');
-    return { ok: true, mtimeMs: 0 };
-  });
-});
-afterAll(() => writeSpy.mockRestore());
-
-const SLUG = '1218-fo-01-move-scripts-lib-into-src-main-lib';
-const FIXTURE_PATH = path.join(__dirname, 'fixtures', `${SLUG}.log`);
-const LANDED_COMMIT = 'd1edf15be50a783f08a55e6de328780f0feb3cb0';
-
-function makeTmpDir() {
-  return fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-landed-commit-outranks-test-'));
-}
-
-function rmdir(dir) {
-  try { fs.rmSync(dir, { recursive: true, force: true }); } catch { /* */ }
-}
-
-function makeTmpGitRepo() {
-  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-landed-commit-outranks-repo-'));
-  execFileSync('git', ['init', '-q'], { cwd: dir });
-  execFileSync('git', ['config', 'user.email', 'test@example.com'], { cwd: dir });
-  execFileSync('git', ['config', 'user.name', 'Test'], { cwd: dir });
-  fs.writeFileSync(path.join(dir, 'README.md'), 'init\n');
-  execFileSync('git', ['add', 'README.md'], { cwd: dir });
-  execFileSync('git', ['commit', '-q', '-m', 'init'], { cwd: dir });
-  return dir;
-}
-
-function writePrd(dir, slug) {
-  const text = '---\ntitle: Test PRD\ncwd: /tmp\nestimateMinutes: 30\n---\n# Move scripts/lib into src/main/lib\n';
-  fs.writeFileSync(path.join(dir, `${slug}.md`), text);
-  return path.join(dir, `${slug}.md`);
-}
-
-/**
- * Load the real fixture log, optionally stripping the trailing
- * `SCHEDULER_VERDICT: FAIL ...` line from the final `{"type":"result",...}`
- * event's `result` field so the sentinel scan sees no sentinel at all.
- */
-function loadFixture({ stripFailSentinel }) {
-  const raw = fs.readFileSync(FIXTURE_PATH, 'utf8');
-  if (!stripFailSentinel) return raw;
-  const lines = raw.split('\n');
-  const idx = lines.findIndex((l) => l.includes('"type":"result"'));
-  assert.ok(idx !== -1, 'fixture must contain a type:result line');
-  const obj = JSON.parse(lines[idx]);
-  assert.ok(/SCHEDULER_VERDICT:\s*FAIL/.test(obj.result), 'fixture result must carry the real FAIL sentinel before stripping');
-  obj.result = obj.result.replace(/\n?SCHEDULER_VERDICT:\s*FAIL[^\n]*$/, '');
-  assert.ok(!/SCHEDULER_VERDICT/.test(obj.result), 'sentinel must be fully removed from the stripped variant');
-  lines[idx] = JSON.stringify(obj);
-  return lines.join('\n');
-}
-
-function writeLog(dir, slug, text) {
-  fs.writeFileSync(path.join(dir, `${slug}.log`), text);
-}
-
-// (a) exit 0 + commit landed THIS run + no sentinel at all → clean, demoted to annotation.
-test('exit 0 + landed commit this run + no FAIL sentinel → clean with annotation (1218-fo-01)', async () => {
-  const tmp = makeTmpDir();
-  const repo = makeTmpGitRepo();
-  try {
-    writeLog(tmp, SLUG, loadFixture({ stripFailSentinel: true }));
-    const prdPath = writePrd(tmp, SLUG);
-    const verdict = await verifyRun({
-      runDir: tmp,
-      prdPath,
-      queueEntry: { slug: SLUG, status: 'running', cwd: repo },
-      allJobs: [],
-      committedDuringRun: false,
-      jobLandedCommitThisRun: LANDED_COMMIT,
-      exitCode: 0,
-    });
-    assert.equal(verdict.verdict, 'clean', `expected clean, got ${verdict.verdict}: ${verdict.reason}`);
-    assert.equal(verdict.downgradeTo, null);
-    const sidecar = JSON.parse(fs.readFileSync(path.join(tmp, `${SLUG}.verdicts.json`), 'utf8'));
-    assert.ok(Array.isArray(sidecar.annotations) && sidecar.annotations.length > 0, 'the demoted is_error hit must be recorded as an annotation');
-    assert.ok(sidecar.annotations.some((a) => a.verdict === 'transcript_errors'), 'annotation must carry the original transcript_errors verdict');
-  } finally { rmdir(tmp); rmdir(repo); }
-});
-
-// (b) the same (stripped-sentinel) log, but exitCode !== 0 → keeps today's behavior.
-test('non-zero exit → transcript_errors (evidence never applies to a failing run)', async () => {
-  const tmp = makeTmpDir();
-  const repo = makeTmpGitRepo();
-  try {
-    writeLog(tmp, SLUG, loadFixture({ stripFailSentinel: true }));
-    const prdPath = writePrd(tmp, SLUG);
-    const verdict = await verifyRun({
-      runDir: tmp,
-      prdPath,
-      queueEntry: { slug: SLUG, status: 'running', cwd: repo },
-      allJobs: [],
-      committedDuringRun: false,
-      jobLandedCommitThisRun: LANDED_COMMIT,
-      exitCode: 1,
-    });
-    assert.equal(verdict.verdict, 'transcript_errors', `expected transcript_errors, got ${verdict.verdict}: ${verdict.reason}`);
-    assert.equal(verdict.downgradeTo, 'needs_review');
-  } finally { rmdir(tmp); rmdir(repo); }
-});
-
-// (c) the pristine, byte-for-byte-unmodified fixture — its own real FAIL sentinel — keeps today's behavior.
-test('explicit FAIL sentinel (the real, unmodified fixture) → transcript_errors byte-for-byte', async () => {
-  const tmp = makeTmpDir();
-  const repo = makeTmpGitRepo();
-  try {
-    writeLog(tmp, SLUG, loadFixture({ stripFailSentinel: false }));
-    const prdPath = writePrd(tmp, SLUG);
-    const verdict = await verifyRun({
-      runDir: tmp,
-      prdPath,
-      queueEntry: { slug: SLUG, status: 'running', cwd: repo },
-      allJobs: [],
-      committedDuringRun: false,
-      jobLandedCommitThisRun: LANDED_COMMIT,
-      exitCode: 0,
-    });
-    assert.equal(verdict.verdict, 'transcript_errors', `expected transcript_errors, got ${verdict.verdict}: ${verdict.reason}`);
-    assert.equal(verdict.downgradeTo, 'needs_review');
-    assert.equal(verdict.sentinel, 'fail');
-  } finally { rmdir(tmp); rmdir(repo); }
-});
-
-// (d) no landed-commit evidence at all → transcript_errors (evidence gate requires attribution).
-test('no landed commit evidence → transcript_errors (evidence gate requires attribution)', async () => {
-  const tmp = makeTmpDir();
-  const repo = makeTmpGitRepo();
-  try {
-    writeLog(tmp, SLUG, loadFixture({ stripFailSentinel: true }));
-    const prdPath = writePrd(tmp, SLUG);
-    const verdict = await verifyRun({
-      runDir: tmp,
-      prdPath,
-      queueEntry: { slug: SLUG, status: 'running', cwd: repo },
-      allJobs: [],
-      committedDuringRun: false,
-      jobLandedCommitThisRun: null,
-      exitCode: 0,
-    });
-    assert.equal(verdict.verdict, 'transcript_errors', `expected transcript_errors, got ${verdict.verdict}: ${verdict.reason}`);
-    assert.equal(verdict.downgradeTo, 'needs_review');
-  } finally { rmdir(tmp); rmdir(repo); }
-});

package/src/main/__tests__/runVerify-policy-denial.test.cjs

@@ -1,89 +0,0 @@
-/**
- * runVerify-policy-denial.test.cjs — a PreToolUse hook denial (`Blocked: ...`
- * tool_result) must not downgrade an otherwise-clean run, but a genuine
- * late-run failure must still downgrade normally.
- *
- * Run: timeout 120 npx vitest run src/main/__tests__/runVerify-policy-denial.test.cjs
- */
-
-import { test } from 'vitest';
-const assert = require('node:assert/strict');
-const os = require('node:os');
-const fs = require('node:fs');
-const path = require('node:path');
-const { verifyRun } = require('../runVerify.cjs');
-
-function makeTmpDir() {
-  return fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-policy-denial-test-'));
-}
-
-function rmdir(dir) {
-  try { fs.rmSync(dir, { recursive: true, force: true }); } catch { /* */ }
-}
-
-function writeLog(dir, slug, events) {
-  const lines = ['[scheduler] starting ' + slug + ' at 2026-09-02T00:00:00.000Z'];
-  for (const ev of events) lines.push(JSON.stringify(ev));
-  lines.push('[scheduler] exit code=0 (raw code=0 signal=null) duration=47s');
-  fs.writeFileSync(path.join(dir, `${slug}.log`), lines.join('\n') + '\n');
-}
-
-function writePrd(dir, slug, body) {
-  const text = `---\ntitle: Test PRD\ncwd: /tmp\nestimateMinutes: 30\n---\n${body}\n`;
-  fs.writeFileSync(path.join(dir, `${slug}.md`), text);
-  return path.join(dir, `${slug}.md`);
-}
-
-test('PreToolUse policy denial (Blocked: ...) in final 20% → clean, surfaced as annotation', async () => {
-  const tmp = makeTmpDir();
-  try {
-    const slug = '1106-policy-denial-exempt';
-    const events = [];
-    for (let k = 0; k < 8; k++) {
-      events.push({ type: 'assistant', message: { role: 'assistant', content: [
-        { type: 'tool_use', id: `t${k}`, name: 'Read', input: { description: `read ${k}` } }] } });
-      events.push({ type: 'user', message: { role: 'user', content: [
-        { type: 'tool_result', tool_use_id: `t${k}`, content: 'ok', is_error: false }] } });
-    }
-    events.push({ type: 'assistant', message: { role: 'assistant', content: [
-      { type: 'tool_use', id: 'tgit', name: 'Bash', input: { command: 'git stash', description: 'stash before checkout' } }] } });
-    events.push({ type: 'user', message: { role: 'user', content: [
-      { type: 'tool_result', tool_use_id: 'tgit',
-        content: 'Blocked: `stash` in what this hook believes is a SHARED working tree (/home/bilko/Projects/session-manager).',
-        is_error: true }] } });
-    events.push({ type: 'result', subtype: 'success', result: 'All acceptance criteria verified.\nSCHEDULER_VERDICT: PASS' });
-
-    writeLog(tmp, slug, events);
-    const prdPath = writePrd(tmp, slug, '# Policy denial exempt');
-    const verdict = await verifyRun({ runDir: tmp, prdPath, queueEntry: { slug, status: 'running' }, allJobs: [], committedDuringRun: true });
-    assert.equal(verdict.verdict, 'clean', `policy denial must not flag, got ${verdict.verdict}: ${verdict.reason}`);
-    assert.ok(
-      Array.isArray(verdict.annotations) && verdict.annotations.some((a) => a.verdict === 'policy_denial'),
-      `expected a policy_denial annotation, got ${JSON.stringify(verdict.annotations)}`,
-    );
-  } finally { rmdir(tmp); }
-});
-
-test('genuine late-run failure (not a policy denial) still downgrades to transcript_errors', async () => {
-  const tmp = makeTmpDir();
-  try {
-    const slug = '1106-genuine-failure-still-flags';
-    const events = [];
-    for (let k = 0; k < 8; k++) {
-      events.push({ type: 'assistant', message: { role: 'assistant', content: [
-        { type: 'tool_use', id: `t${k}`, name: 'Read', input: { description: `read ${k}` } }] } });
-      events.push({ type: 'user', message: { role: 'user', content: [
-        { type: 'tool_result', tool_use_id: `t${k}`, content: 'ok', is_error: false }] } });
-    }
-    events.push({ type: 'assistant', message: { role: 'assistant', content: [
-      { type: 'tool_use', id: 'ttest', name: 'Bash', input: { command: 'npm test', description: 'run tests' } }] } });
-    events.push({ type: 'user', message: { role: 'user', content: [
-      { type: 'tool_result', tool_use_id: 'ttest', content: 'FAILED: 1 test failed', is_error: true }] } });
-    events.push({ type: 'result', subtype: 'success', result: 'All acceptance criteria verified.' });
-
-    writeLog(tmp, slug, events);
-    const prdPath = writePrd(tmp, slug, '# Genuine failure still flags');
-    const verdict = await verifyRun({ runDir: tmp, prdPath, queueEntry: { slug, status: 'running' }, allJobs: [], committedDuringRun: true });
-    assert.equal(verdict.verdict, 'transcript_errors', `genuine failure must still flag, got ${verdict.verdict}: ${verdict.reason}`);
-  } finally { rmdir(tmp); }
-});

package/src/main/__tests__/runVerify-transcript-commit-evidence.test.cjs

@@ -1,225 +0,0 @@
-/**
- * runVerify-transcript-commit-evidence.test.cjs — regression tests for the
- * transcript-commit-evidence fallback added to runVerify.cjs.
- *
- * Root cause (99-fix-e2e-needs-review-test, 2026-08-07): a job whose cwd is
- * not a git repo (e.g. an investigation fix-plan that inherited cwd: /tmp)
- * can never produce committedDuringRun=true — gitHead()/committedInWindow()
- * both fail closed there — so the PASS+commit sentinel override could never
- * fire even when the harness's own transcript recorded a real commit. This
- * file asserts the fallback fires ONLY when both conditions hold: the cwd is
- * genuinely not a git repo, AND the transcript itself carries harness-emitted
- * commit evidence (never inferred from prose the model could fabricate).
- *
- * Run standalone: timeout 120 npx vitest run src/main/__tests__/runVerify-transcript-commit-evidence.test.cjs
- */
-
-import { test, beforeAll, afterAll, vi } from 'vitest';
-const assert = require('node:assert/strict');
-const os = require('node:os');
-const fs = require('node:fs');
-const path = require('node:path');
-const { execFileSync } = require('node:child_process');
-const { verifyRun } = require('../runVerify.cjs');
-
-// Sidecar goes through config.writeJsonSync, whose allow-list rejects os.tmpdir() fixtures.
-let writeSpy;
-beforeAll(() => {
-  writeSpy = vi.spyOn(require('../config.cjs'), 'writeJsonSync').mockImplementation((abs, data) => {
-    require('node:fs').writeFileSync(abs, JSON.stringify(data, null, 2) + '\n');
-    return { ok: true, mtimeMs: 0 };
-  });
-});
-afterAll(() => writeSpy.mockRestore());
-
-// ─── helpers (mirrors runVerify.test.cjs) ─────────────────────────────────────
-
-function makeTmpDir() {
-  return fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-commit-evidence-test-'));
-}
-
-function rmdir(dir) {
-  try { fs.rmSync(dir, { recursive: true, force: true }); } catch { /* */ }
-}
-
-function writeLog(dir, slug, events) {
-  const lines = ['[scheduler] starting ' + slug + ' at 2026-08-07T02:25:31.453Z'];
-  for (const ev of events) lines.push(JSON.stringify(ev));
-  lines.push('[scheduler] exit code=0 (raw code=0 signal=null) duration=47s');
-  fs.writeFileSync(path.join(dir, `${slug}.log`), lines.join('\n') + '\n');
-}
-
-function writePrd(dir, slug, body) {
-  const text = `---\ntitle: Test PRD\ncwd: /tmp\nestimateMinutes: 30\n---\n${body}\n`;
-  fs.writeFileSync(path.join(dir, `${slug}.md`), text);
-  return path.join(dir, `${slug}.md`);
-}
-
-function makeTmpGitRepo() {
-  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-commit-evidence-repo-'));
-  execFileSync('git', ['init', '-q'], { cwd: dir });
-  execFileSync('git', ['config', 'user.email', 'test@example.com'], { cwd: dir });
-  execFileSync('git', ['config', 'user.name', 'Test'], { cwd: dir });
-  fs.writeFileSync(path.join(dir, 'README.md'), 'init\n');
-  execFileSync('git', ['add', 'README.md'], { cwd: dir });
-  execFileSync('git', ['commit', '-q', '-m', 'init'], { cwd: dir });
-  return dir;
-}
-
-/**
- * Build a run log with:
- *   - padding events so a late is_error lands in the final 20%
- *   - a late is_error tool_result (main-agent, no self-recovery)
- *   - optionally a harness-emitted commit-evidence line
- *   - a SCHEDULER_VERDICT: PASS result event
- */
-function buildRunEvents({ includeCommitEvidence, evidenceForm = 'vcs_state_changed' }) {
-  const events = [];
-  for (let k = 0; k < 8; k++) {
-    events.push({ type: 'assistant', message: { role: 'assistant', content: [
-      { type: 'tool_use', id: `t${k}`, name: 'Read', input: { description: `read ${k}` } }] } });
-    events.push({ type: 'user', message: { role: 'user', content: [
-      { type: 'tool_result', tool_use_id: `t${k}`, content: 'ok', is_error: false }] } });
-  }
-  events.push({ type: 'assistant', message: { role: 'assistant', content: [
-    { type: 'tool_use', id: 'tbad', name: 'Agent', input: { description: 'run code review agent' } }] } });
-  events.push({ type: 'user', message: { role: 'user', content: [
-    { type: 'tool_result', tool_use_id: 'tbad',
-      content: "Agent type 'code-reviewer' not found. Available agents: architect, …", is_error: true }] } });
-
-  if (includeCommitEvidence) {
-    if (evidenceForm === 'vcs_state_changed') {
-      events.push({ type: 'system', subtype: 'vcs_state_changed', kind: 'commit', cwd: '/tmp' });
-    } else {
-      events.push({
-        type: 'user',
-        message: { role: 'user', content: [
-          { type: 'tool_result', tool_use_id: 'tcommit', content: 'committed', is_error: false }] },
-        tool_use_result: { gitOperation: { commit: { sha: '6ecbd40', kind: 'committed' } } },
-      });
-    }
-  }
-
-  events.push({ type: 'result', subtype: 'success', result: 'All acceptance criteria verified.\nSCHEDULER_VERDICT: PASS' });
-  return events;
-}
-
-// ─── (1) regression case: non-git cwd + transcript evidence → clean ──────────
-
-test('PASS + late is_error + transcript commit evidence + non-git cwd → clean (fallback fires)', async () => {
-  const tmp = makeTmpDir();
-  try {
-    const slug = '99-fix-e2e-needs-review-test';
-    writeLog(tmp, slug, buildRunEvents({ includeCommitEvidence: true }));
-    const prdPath = writePrd(tmp, slug, '# Fix scheduler skip guard');
-    const nonGitCwd = fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-nongit-cwd-'));
-    try {
-      const verdict = await verifyRun({
-        runDir: tmp,
-        prdPath,
-        queueEntry: { slug, status: 'running', cwd: nonGitCwd },
-        allJobs: [],
-        committedDuringRun: false,
-      });
-      assert.equal(verdict.verdict, 'clean', `expected clean via transcript fallback, got ${verdict.verdict}: ${verdict.reason}`);
-      assert.equal(verdict.downgradeTo, null);
-      const sidecar = JSON.parse(fs.readFileSync(path.join(tmp, `${slug}.verdicts.json`), 'utf8'));
-      assert.equal(sidecar.commitEvidenceSource, 'transcript', 'sidecar should record the fallback source');
-    } finally { rmdir(nonGitCwd); }
-  } finally { rmdir(tmp); }
-});
-
-test('PASS + late is_error + gitOperation.commit evidence + non-git cwd → clean (fallback fires)', async () => {
-  const tmp = makeTmpDir();
-  try {
-    const slug = '99-fix-e2e-needs-review-test-alt';
-    writeLog(tmp, slug, buildRunEvents({ includeCommitEvidence: true, evidenceForm: 'gitOperation' }));
-    const prdPath = writePrd(tmp, slug, '# Fix scheduler skip guard');
-    const nonGitCwd = fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-nongit-cwd-'));
-    try {
-      const verdict = await verifyRun({
-        runDir: tmp,
-        prdPath,
-        queueEntry: { slug, status: 'running', cwd: nonGitCwd },
-        allJobs: [],
-        committedDuringRun: false,
-      });
-      assert.equal(verdict.verdict, 'clean', `expected clean via transcript fallback, got ${verdict.verdict}: ${verdict.reason}`);
-    } finally { rmdir(nonGitCwd); }
-  } finally { rmdir(tmp); }
-});
-
-// ─── (2) no evidence lines → fallback must not fire, still transcript_errors ──
-
-test('PASS + late is_error + NO transcript evidence + non-git cwd → transcript_errors (no fallback)', async () => {
-  const tmp = makeTmpDir();
-  try {
-    const slug = '99-fix-no-evidence';
-    writeLog(tmp, slug, buildRunEvents({ includeCommitEvidence: false }));
-    const prdPath = writePrd(tmp, slug, '# Fix scheduler skip guard');
-    const nonGitCwd = fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-nongit-cwd-'));
-    try {
-      const verdict = await verifyRun({
-        runDir: tmp,
-        prdPath,
-        queueEntry: { slug, status: 'running', cwd: nonGitCwd },
-        allJobs: [],
-        committedDuringRun: false,
-      });
-      assert.equal(verdict.verdict, 'transcript_errors', `expected transcript_errors without evidence, got ${verdict.verdict}: ${verdict.reason}`);
-      assert.equal(verdict.downgradeTo, 'needs_review');
-    } finally { rmdir(nonGitCwd); }
-  } finally { rmdir(tmp); }
-});
-
-// ─── (3) git-repo cwd + evidence + committedDuringRun:false → fallback must not apply ──
-
-test('PASS + late is_error + transcript evidence + GIT-REPO cwd + committedDuringRun:false → transcript_errors (false-PASS guard)', async () => {
-  const tmp = makeTmpDir();
-  const repo = makeTmpGitRepo();
-  try {
-    const slug = '99-fix-git-repo-cwd';
-    writeLog(tmp, slug, buildRunEvents({ includeCommitEvidence: true }));
-    const prdPath = writePrd(tmp, slug, '# Fix scheduler skip guard');
-    const verdict = await verifyRun({
-      runDir: tmp,
-      prdPath,
-      queueEntry: { slug, status: 'running', cwd: repo },
-      allJobs: [],
-      committedDuringRun: false,
-    });
-    assert.equal(verdict.verdict, 'transcript_errors', `fallback must not apply when git ancestry is observable, got ${verdict.verdict}: ${verdict.reason}`);
-    assert.equal(verdict.downgradeTo, 'needs_review');
-  } finally { rmdir(tmp); rmdir(repo); }
-});
-
-// ─── (4) PASS + transcript evidence + non-git cwd must not be pass_no_commit ──
-
-test('PASS (no error) + transcript evidence + non-git cwd → clean, not pass_no_commit', async () => {
-  const tmp = makeTmpDir();
-  try {
-    const slug = '99-fix-clean-pass';
-    const events = [
-      { type: 'assistant', message: { role: 'assistant', content: [
-        { type: 'tool_use', id: 't0', name: 'Bash', input: { description: 'run tests' } }] } },
-      { type: 'user', message: { role: 'user', content: [
-        { type: 'tool_result', tool_use_id: 't0', content: '535 passed', is_error: false }] } },
-      { type: 'system', subtype: 'vcs_state_changed', kind: 'commit', cwd: '/tmp' },
-      { type: 'result', subtype: 'success', result: 'All acceptance criteria verified.\nSCHEDULER_VERDICT: PASS' },
-    ];
-    writeLog(tmp, slug, events);
-    const prdPath = writePrd(tmp, slug, '# Fix scheduler skip guard');
-    const nonGitCwd = fs.mkdtempSync(path.join(os.tmpdir(), 'run-verify-nongit-cwd-'));
-    try {
-      const verdict = await verifyRun({
-        runDir: tmp,
-        prdPath,
-        queueEntry: { slug, status: 'running', cwd: nonGitCwd },
-        allJobs: [],
-        committedDuringRun: false,
-      });
-      assert.equal(verdict.verdict, 'clean', `expected clean, got ${verdict.verdict}: ${verdict.reason}`);
-      assert.notEqual(verdict.verdict, 'pass_no_commit');
-    } finally { rmdir(nonGitCwd); }
-  } finally { rmdir(tmp); }
-});