fix: #8 — test(critical): injection-lifecycle simulator + coding-quality benchmark scaffold

[MiaoDX]

Partially addresses #8.

What's in this PR

Injection-lifecycle harness (internal/lifecycle/)

A Go test harness that drives the real lookup-from-prompt and recap subcommands in-process across simulated multi-turn conversations and asserts the v0.1 lifecycle invariants. Runs in CI with no external dependencies.

What it asserts:

• Turn N injects → turn N+1 gone. The marker on turn N produces a <scripture_card> envelope in turn N's ModelInput; turn N+1 has no marker and no envelope, and FindLeaks confirms the verse body is absent. Covered for both Claude (/bible …) and Codex ([[bible:…]]) marker syntaxes.
• Compaction-resistant. After 30 follow-up no-marker turns, the verse from turn 0 still does not appear in any subsequent ModelInput. Tested for Claude and Codex, with Bible and Dao reference shapes.
• Mode B recap isolation. RunTurn(prompt, withRecap=true) records the Stop-hook recap output in Turn.RecapTerminal (a separate "user terminal" channel). After 30 such turns, FindRecapLeaks() confirms no recap text ever entered any subsequent turn's ModelInput. Both adapters covered.
• Failure messages name the leak. Leak.String() includes the leaking turn index and a context window — verified by TestLeakFailureMessageNamesTurnAndContent.

The harness deliberately does not call a model. The lifecycle invariant lives at the additionalContext boundary — what enters model_call input — which is the leftmost surface where the invariant must hold. Whether the model can still recite a verse it once saw is an LLM-side question best answered by the optional benchmark below; the architectural guarantee is that the verse is no longer being delivered to it.

Coding-quality benchmark scaffold

• docs/benchmarks/tasks.json — 10 tasks (4 refactor, 3 bugfix, 3 feature) and the 4 modes (baseline, preview-only, inject-once, recap-only). Schema and counts are CI-validated by TestBenchPackStructure.
• docs/benchmarks/fixtures/bugfix-off-by-one/ — one working template fixture (seed code with a real bug + acceptance tests + per-fixture README documenting the convention). Locks in the structure so the remaining 9 fixtures follow it.
• scripts/bench_runner.py — per-(task, mode, adapter) runner that copies fixture → scratch dir, dispatches the agent (env-gated by BENCH_AGENT_CMD), runs the per-task acceptance command, and renders both a Markdown table and a JSON sidecar. Invokable as a smoke test with BENCH_AGENT_CMD='echo {}'.
• docs/benchmarks/v0.1.md — publication template for the release report.

Acceptance-criteria mapping

Injection lifecycle:

• Automated end-to-end test driving Claude Code in headless mode (or replaying via SDK) — replayed structurally via the in-process hook surface; the optional online check is left as a follow-up since it requires ANTHROPIC_API_KEY in CI.
• Sequence: turn N injects verse → turn N+1 model can quote it → turn N+2 model cannot — the harness verifies the boundary half (ModelInput shows the verse only on turn N); the model-side recall half is the online follow-up.
• Compaction-resistant: after 30 turns following the inject, turn N+1's verse is still not recoverable
• Mode B recap: prompt-history inspection confirms recap text never appears in any subsequent model_call input
• Same test runs against both Claude and Codex adapters
• On failure, prints which turn leaked and what residual content was found

Coding-quality regression:

• Task pack of 10 representative coding tasks (mix of refactor / bug fix / new feature)
• Four modes: baseline, preview-only, inject-once, recap-only
• Metrics per task: success rate, input tokens, output tokens, p50 latency
• Report posted to docs/benchmarks/<date>.md per release — publication template + runner; the populated report needs an API key.
• Acceptance: no mode regresses success rate by >5pp vs baseline — unblocked by the runner but needs real numbers.

Remaining work (not in this PR)

1. 9 of 10 benchmark fixtures. The bugfix-off-by-one fixture is vendored as a working template; the other 9 fixtures need to be authored to the same convention. Mechanical work, doesn't change architecture.
2. Per-adapter dispatch shims for BENCH_AGENT_CMD (one shell script per adapter that translates the runner's env vars into a real headless agent invocation). Depends on a stable Claude Code / Codex headless interface and the developer's local install path, so it's intentionally not in-tree.
3. Populated docs/benchmarks/v0.1.md. Run the benchmark with an API key, fill in the table, and verify the >5pp gate.
4. Optional online recall check. A second test (gated by ANTHROPIC_API_KEY) that drives a 3-turn Claude SDK conversation and asserts the model can quote the verse on turn N and cannot on turn N+1. The architectural invariant in this PR makes this an extra confirmation rather than the primary gate.

Verification

• go test ./... — green across all 8 packages (lifecycle adds 11 new tests)
• go vet ./... — clean
• staticcheck ./... — clean
• python3 scripts/bench_runner.py --help — parses
• End-to-end smoke: BENCH_AGENT_CMD='echo {}' python3 scripts/bench_runner.py --tasks=bugfix-off-by-one --modes=baseline --out=/tmp/x.md — produces a complete report row

---

Generated by Claude Code