PR 28651 real command verification
Baseline: github/main 82b10d546c
Original PR head: 290c8924b05f87de8265c2918a4f95c840f614b4
Verified fix: d8f62ccae6824d107c1fd5cceefbe1e9a9c3d662

All cached-test commands and checks below ran on task-owned Blacksmith workers
from the real checkout source with vp 1.4.0 and installed repo dependencies.
No mocked vp executable or summary was used as the current run's result.
A file at VITE_CACHE_PATH and a corrupt stale sibling exercised filesystem failures.

BEFORE (github/main):
VITE_CACHE_PATH=/tmp/m057-main/cache-file VP_REMOTE_CACHE=off \
GITHUB_STEP_SUMMARY=/tmp/m057-main/step.md node scripts/test-cached.ts \
  --report /tmp/m057-main/report.json --verify-hits 1 scripts/lib/sharded-tests.test.ts
Cached tests: 1 files, 1 per task, concurrency 1
Failed to initialize cache: Not a directory (os error 20)
Error: ENOTDIR: not a directory, scandir '/tmp/m057-main/cache-file'
EXIT=1
No report or step summary was written.

ORIGINAL PR HEAD, same file-at-cache-root scenario:
Failed to initialize cache: Not a directory (os error 20)
No task summary from this run in /tmp/m057-pr/cache-file
EXIT=1
Report: exitCode=1, summary=null, verification=[]
Step summary: 1 files selected; 0 hits; 0 cached results re-executed.

NEIGHBORING FAILURE on original PR head:
Real cached-test command selected scripts/lib/sharded-tests.test.ts.
Clean cache: test passed, CLEAN_EXIT=0, report included its own run marker.
After adding cache/old-broken/last-summary.json containing invalid JSON:
Test passed, but: No task summary from this run in /tmp/m057-pr/cache
CORRUPT_EXIT=1; report.exitCode=0, report.summary=null.
Existing summary regression case extended with corrupt sibling:
1 failed: AssertionError: Target cannot be null or undefined.

AFTER FIX:
Each candidate summary is read and parsed independently; unreadable or corrupt
siblings are skipped while searching for this run's marker.
vp test run scripts/lib/cached-tests.test.ts: 8 passed.
Includes real-vp cache reuse, invalidation, uncached paths and failed-test coverage.
Same real command with corrupt sibling retained: FIXED_EXIT=0, own summary found.
Repeat with --verify-hits 1: HIT_EXIT=0, hits=1, verification[0].exitCode=0.
Own marker: node -e "" b12b68c1a4083e731cadb4b29f6c9e81
Step summary: 1 cached results re-executed without cache; 0 verification failures.
File-at-cache-root scenario repeated after fix:
FINAL_MISSING_EXIT=1, report.exitCode=1, summary=null, verification=[].
Report and step summary both present; no stack trace.

SCOPED CHECKS:
blacksmith-check run ... 'vp run check:changed 8760d063e903406c8e8f695b0852c0a15fc1e00f'
format: pass
lint: pass, 0 warnings, 0 errors
knip 1/2 and 2/2: pass
typecheck (6 affected packages): pass
test (2 files): pass, 9 tests

REAL CLIENT / PROVIDER HEALTH:
Installed with vp i. One background dev server at a time, main stopped by captured
process group before PR server started. PR server restarted after the fix.
PR dev server: serverPort=13773, webPort=5733, isolated .t3 home, automatic seed off.
Playwright from repo dependencies drove /usr/bin/google-chrome.
Added /home/user/m057-project through Add project > Local folder.
Selected Codex gpt-6.1-sol in the model picker and submitted a real turn.
Tool read README.md and git branch; returned '# audit m057 verification' and 'master'.
After restart, same thread and project persisted. Follow-up real Codex turn read
README.md and git status and returned 'workspace clean: audit m057 verification'.
No workspace files changed. Screenshot and recording show the actual web client.
This is app/provider smoke evidence; the changed behavior is the cached-test CLI.

NOT VERIFIED:
Hosted CI cache-initialization failure and remote-cache hits.
Native desktop, iOS/Android, macOS and Windows execution.
Concurrent writers may overwrite the single vp summary; a mismatched run must be
rejected rather than reporting another run's tasks. Existing run-identity test passes.