test(bench): vendor 9 remaining benchmark fixtures for issue #8

[MiaoDX]

Continues #8 (supersedes #19). Picks up the largest mechanical bullet from the remaining-work checklist on #8: vendor the 9 fixtures referenced by docs/benchmarks/tasks.json whose absence was tracked in the issue comment.

Branch name note: this is the issue-#8 follow-up branch; the previous attempt on claude-issue-8 had a tangled history with PR #17's squash-merge and produced a content-conflict, so this PR is on claude-issue-8-fixtures (single commit off main).

What's in this PR

Nine fixtures, one convention

Each follows the bugfix-off-by-one template (PR #17): README.md describing the task in PR-description form, one source file with the seed code, and a test_*.py acceptance suite that the runner grades with pytest -q.

id	type	hook the seed exposes
refactor-cron	refactor	parsing inlined in schedule_job (extract parse_cron)
refactor-config-merge	refactor	nested if-cascade for 3-layer merge (introduce deep_merge)
refactor-csv-rows	refactor	row[0] / row[1] / row[2] indexing (switch to namedtuple)
refactor-rename	refactor	function do + variable tmp (rename, keep public API)
bugfix-utf8-truncate	bugfix	s.encode()[:n] splits multi-byte codepoints
bugfix-race-counter	bugfix	explicit read-modify-write with time.sleep(0) between read and write
feature-rate-limit	feature	TokenBucket + rate_limit decorator stubs raising NotImplementedError
feature-flatten-json	feature	flatten_json(obj, sep) stub
feature-cli-flag	feature	working text CLI; needs opt-in --json flag + render_json helper

Seed-code grading is correct

Each fixture's seed already fails its acceptance suite, so a no-op agent run grades as a miss. Confirmed locally:

• Refactor seeds pass behavioral tests but fail the structural assertions (e.g. parse_cron exists, no row[0..2] substrings remain, no module-level do / no \btmp\b token, deep_merge exists).
• Bugfix and feature seeds fail the relevant behavior tests directly (UnicodeDecodeError on multi-byte split, lost increments under threads, NotImplementedError, AssertionError on missing --json).

The race-counter seed is an explicit non-atomic read-modify-write with time.sleep(0) between the read and the write — necessary because plain self._n += 1 is usually atomic on CPython (GIL) and would not reliably grade as a bug. With the explicit yield, races land deterministically (the test loses ~96% of updates on this machine).

Test infrastructure

• internal/lifecycle/bench_pack_test.go: TestBenchTemplateFixtureExists is replaced with TestBenchFixturesPresent, which iterates every task in tasks.json and asserts the fixture dir contains a README.md, a non-test source .py, and a test_*.py. Adding a task without vendoring the fixture (or vice versa) now fails CI.
• .gitignore: adds __pycache__/ and *.pyc since pytest now runs against the fixtures during local verification.

Acceptance-criteria mapping (issue #8)

This PR closes the "task pack of 10 representative coding tasks" sub-bullet, which was previously satisfied only on paper (10 entries in tasks.json, 1 fixture vendored).

Still outstanding on #8 — and the same items the prior comment flagged:

• Per-adapter BENCH_AGENT_CMD shims (intentionally out-of-tree per fix: #8 — test(critical): injection-lifecycle simulator + coding-quality benchmark scaffold #17's notes)
• Populated docs/benchmarks/v0.1.md (requires ANTHROPIC_API_KEY)
• Optional online recall check (requires ANTHROPIC_API_KEY)

Verification

• go test ./... — green across all 10 packages
• go vet ./... — clean
• python3 -m py_compile on every fixture source — clean
• pytest -q on each seed fixture grades as expected (failures land on the bug / missing functionality the agent is asked to fix)
• BENCH_AGENT_CMD='echo {}' python3 scripts/bench_runner.py --tasks=refactor-cron,bugfix-utf8-truncate --modes=baseline --out=/tmp/x.md — runner produces a complete report row for the new fixtures

---

Generated by Claude Code