# Make test failures concise and directly actionable

## Traceability

- Spec ID: test-ci-diagnostics
- Status: Implemented

## Intent

Keep the dependency-free `node:test` suite while making GitHub Actions failures
easy to locate. Replace the unbounded default TAP stream with capability groups,
compact progress, file-and-line annotations, a job summary, and downloadable
JUnit results. Continue removing assertions that merely freeze private HTML/CSS
spelling when an executable validator already owns the behavior.

## Acceptance Scenarios

- AC-1: CI runs every discovered capability directory, continues after one
  category fails, and exits non-zero after reporting all failed categories.
- AC-2: Passing tests use compact console output; a failing test emits a GitHub
  annotation containing its name, repository-relative file, line, and useful
  error detail.
- AC-3: Every CI matrix job writes one JUnit XML result per capability and
  uploads the files even when tests fail.
- AC-4: GitHub's step summary lists category status, test-file count, and
  duration without requiring a third-party test framework or reporter action.
- AC-5: Local `npm test` remains the direct full-suite command, while
  `npm run test:ci` owns CI-only presentation and artifacts.
- AC-6: Remaining report renderer and task-loop projection tests retain data
  contracts, escaping, accessibility, validator failures, and self-contained
  artifact boundaries but stop pinning private CSS dimensions and redundant
  visible-copy spelling.
- AC-7: Tests that repeat an expensive package build or Git fixture setup split
  reusable mechanics from one representative integration path, while retaining
  the same portability, replacement, CLI, and analysis contracts.
- AC-8: The categorized runner passes test modules as portable paths relative
  to its working directory and loads its custom reporter through a file URL, so
  Windows drive letters are never interpreted as ESM URL schemes.

## Non-goals

- Migrating the existing suite to Vitest or Jest without a missing capability
  that Node 22 cannot provide.
- Publishing test infrastructure in the npm package.
- Adding coverage percentage gates or changing product behavior.
- Editing the concurrently modified Inspector and commit-session test slices.

## Plan and Tasks

1. Add test-only CI infrastructure under `test/support`: discover capability
   directories, run them sequentially with compact GitHub and JUnit reporters,
   and aggregate failures.
2. Add focused behavior tests for discovery, annotation escaping, and a real
   failing child test.
3. Use the CI runner in `.github/workflows/ci.yml` and upload `test-results` on
   every matrix job.
4. Remove renderer and task-loop projection assertions that duplicate
   validators or freeze layout/copy literals, leaving independent semantic and
   security oracles intact.
5. Verify focused tests, a successful CI-mode run, the full local suite, package
   contents, and staged/unstaged boundaries.
6. Use JUnit timings to remove repeated expensive setup from the slowest tests
   without weakening cross-platform, security, or end-to-end boundaries.
7. Normalize test and reporter module arguments for the Windows Node test
   runner, and cover the drive-letter projection with a platform-independent
   behavior test.

## Test and Review Evidence

- AC-1..AC-4/AC-8: the CI reporting behavior tests pass 5/5 on local Node 24,
  including a real failed-first/success-second capability run and a Windows
  drive-letter path projection. `npm run test:ci` passes all 8 capability
  groups and writes 8 JUnit XML files.
- AC-5: `npm test` passes 1,310/1,310 after adding the Windows runner
  regression test.
- AC-6: the renderer tests pass 33/33 and the task-loop projection tests pass
  84/84. These files contain 60 fewer regex assertions: 48 private HTML/CSS or
  visible-copy checks and 12 redundant Markdown spelling checks.
- AC-7: JUnit profiling identified three repeated setup costs. The Codex plugin
  artifact test now performs one full build and tests atomic replacement in
  isolation; CLI help validates all 53 terminal routes structurally and runs 8
  representative guarded owners; both AI census surfaces share one 7-commit
  repository fixture. On the local Node 24 CI path, the affected slow cases
  moved from 44.3s to about 19.5s, 16.8s to about 4.2s, and 21.3s per file to
  about 1.5s respectively.
- Documentation: the focused doc-link graph suite passes 6/6 after regenerating
  the canonical graph.
- Package: `npm run pack:verify` passes with 507 npm entries and 529 runtime zip
  entries. `git diff --check` is clean.
- Risk: concurrent full-suite runs can exhaust Git fixture resources and create
  unrelated timeout failures. Run final suites serially and distinguish an
  environment timeout from an assertion regression.