test: decide and wire a testing strategy for grimoire's executable code (#34)

[Jartans-Familiar]

Closes #34.

Stacked on #51 (agent/loom/1ea9712f) -- the Node test file lives beside
find-duplicate-comments.js at its post-split location
(plugins/praxis/skills/comment-hygiene/), so this PR's base is that
branch rather than main. Merge #51 first; this one will retarget main
automatically once it does.

Decision (the questions #34 asked)

• Node: node:test + node:assert/strict. Ships with the Node already
  running every hook -- zero new dependency, no package.json/lockfile to
  add. Tests colocate with source as *.test.js, node:test's own
  zero-config discovery pattern (node --test with no path args finds them
  recursively from repo root).
• Python: stdlib unittest, not pytest -- the same reasoning
  .pre-commit-config.yaml already states for skipping ruff: a pyproject.toml
  to hold config that nothing else in the repo needs. Tests live in scripts/
  as test_*.py (unittest discover's default pattern) and load
  generate-codex.py via importlib.util, since its hyphenated filename
  can't be a plain import.
• One strategy or two: two, deliberately -- each stays idiomatic to its
  own runtime's zero-config discovery rather than fighting either tool to
  force a shared filename convention across languages.
• Scope: functions that carry real parsing or bookkeeping logic.
  find-duplicate-comments.js's exported pure functions (prose,
  buildSkipMatcher, findRetoldInDiff, dedupePairs -- seamed for this in
  feat: repo hygiene enforcement, plus praxis compact and skill-activation fixes #33, orphaned since, per Decide a testing strategy for the marketplace's executable code #34's own description) get 19 cases covering the
  three defect classes feat: repo hygiene enforcement, plus praxis compact and skill-activation fixes #33's review rounds caught by hand: the diff
  line-number off-by-one, the unescaped---skip-pattern false-clean, and the
  CRLF false-clean. generate-codex.py's parse_frontmatter, qualify,
  display_name, short_description, basic_string (the TOML escaper) and
  render_agent_toml get 21 cases.
  Out of scope, named rather than silently dropped: the thin hook
  wrappers (check-console-log.js, suggest-compact.js, suggest-skills.js,
  sync-codex-agents.js) -- each is mostly a call into the Claude Code hook
  API, where a fixture would assert little; and generate-codex.py's
  filesystem-walking functions (collect_targets, find_orphans,
  find_divergent_copies, main), which already run end to end against the
  real repo on every make verify --check -- that IS their regression
  coverage, and duplicating it under a mock filesystem would test the mock,
  not the logic.
• Wiring: make test, a new test job in ci.yml (needs
  actions/setup-node alongside the existing setup-python), and both the
  needs: list and the results array of the check aggregator, per the
  hand-sync note that file already carries. CLAUDE.md, CONTRIBUTING.md
  and README.md all updated so make test is part of the stated
  pre-PR/pre-done checklist, not a step only CI knows about.

Demonstrated: red before, green after

Both done locally against this branch, then reverted before committing --
neither seeded defect is in the diff:

Node -- removed the "no newline at end of file" marker guard in
findRetoldInDiff (the exact historical off-by-one #34 names):

✖ keeps line numbers correct across a "no newline at end of file" marker
  AssertionError: Expected values to be strictly equal:
  + 'new/file.js:4'
  - 'new/file.js:3'

Restored -> node --test: 19/19 pass.

Python -- swapped basic_string's escape order (quote before backslash
instead of after), which lets a TOML parser read the output as ending the
string early rather than as an escaped quote:

FAIL: test_escaping_a_backslash_and_a_quote_together_stays_parseable
AssertionError: '"a\\\\\\\\"b"' != '"a\\\\\\"b"'

Restored -> python3 -m unittest discover: 21/21 pass.

Domain-owner clause

Artifact-only. This is test infrastructure and CI wiring for code Loom
already owns test coverage for per the role spec; no skill or agent
content changes.

Manifest / Codex face

praxis bumped 1.5.0 -> 1.5.1 (PATCH: the new test file is
non-behavioural -- it ships as plugin content because it's colocated under
skills/comment-hygiene/, but nothing at runtime reads it). Codex face:
general-reviewer.toml and the plugin manifests regenerated by
scripts/generate-codex.py; nothing under plugins/praxis/agents/ or
codex/ changed content-wise in this PR, so the regen diff is the version
line only.

Verification

make verify   # clean
make test     # 19 Node + 21 Python, all pass
make lint     # clean (lychee-docker/actionlint-docker skip in this sandbox --
              # no Docker socket access; codespell/shellcheck/markdownlint-cli2/
              # zizmor/pre-commit-hooks all pass)

Task JAR-413.

[Jartans-Familiar]

Holding this one. #34 was decided and closed by #50, merged as ee341bd, and the convention now lives in CONTRIBUTING.md under ## Tests. This PR re-answers the same question and lands on the other side of it in four places, so it cannot merge as written:

1. Test location. This colocates (plugins/praxis/skills/comment-hygiene/find-duplicate-comments.test.js, scripts/test_generate_codex.py). The merged convention puts tests in tests/, for two repo-specific reasons: everything under plugins/<name>/ ships to whoever installs that plugin and a test suite is not part of the offer, and a content change there requires a version bump in its plugin.json -- so a colocated test forces a plugin release on every test edit.
2. Discovery. This relies on node --test with no path arguments and unittest discover. The merged convention requires explicit git ls-files lists: node --test reads a directory argument on Node 18 and rejects it on Node 24, and takes a ** glob on Node 24 and rejects it on Node 18, so a file list is the only form that works on both -- which matters because a contributor's Node and CI's Node differ. unittest discover separately wants an __init__.py in every test directory.
3. Hook entry points. Named out of scope here as thin wrappers. The merged convention makes a process contract per shipped hook entry point required (class 2): a hook that throws breaks the user's turn, and no other check in this repository can see it. Those four contracts are in JAR-430: cover the hook scripts and their process contracts #54.
4. The drift finders. find_orphans and find_divergent_copies are out of scope here on the grounds that make verify exercises them. It exercises them against one tree state -- the current, clean one -- which is the state in which they return nothing. Neither the orphan case nor the divergence case is reached, so that is not regression coverage for either. Both are class 1 in the merged convention and are covered in JAR-430: wire make test and cover the two parsing-heavy files #52.

Also: Closes #34 will not do anything, since #50 already closed it.

The gate itself collides outright -- both branches define make test and a test job plus the check aggregator edit, on the same lines.

#52 and #54 implement the merged convention: 139 Node tests and 40 Python tests, the four hook process contracts, and the four find-duplicate-comments.js regressions each confirmed to fail with its fix reverted.

What is worth keeping from here is the coverage, not the wiring. If any of the 40 cases in this branch assert something #52 misses, port those cases onto tests/ and close this. Otherwise close it. The decision is not reopenable from a PR -- if the location or discovery rule is wrong, that is an amendment to CONTRIBUTING.md argued on its own.

[JartanFTW]

resolve conflicts