feat(check): opt-in session-trace archive for multi check

[RobbieMcKinstry]

What

Adds opt-in session-trace capture to multi check, bundled into a single, in-memory .tar.gz for debugging failed (and all other) agent runs.

--trace-archive <PATH> — off by default; also settable as MULTI_CHECKS_TRACE_ARCHIVE or [checks].trace_archive. When set, every check execution's agent session is captured (not just failures).

Why on_event, not cersei's .memory() persistence

The obvious approach — attach a Memory backend so cersei writes its transcript — would miss exactly the runs worth debugging. cersei flushes memory.store() only on the agentic loop's normal exit, but CerseiExecutor cancels its agent the instant a verdict lands and wraps the run in a drop-on-timeout — so successful (cancelled), timed-out, and errored runs all skip the store. Capture is instead via Agent::on_event, which emit() fires synchronously for every event before those early returns, so the trace survives cancellation/timeout/error.

How it fits together

• Executor (executor/trace.rs): a coalescing TraceRecorder (text/thinking deltas joined into per-turn blocks; tool results bounded) serializes each execution into a self-contained NDJSON doc — header line, records, outcome footer — returned in AgentOutcome::trace_jsonl.
• Execution actor: pushes each attempt (retries included) into a shared TraceCollector before signalling completion downstream, so retry traces are captured and collection is race-free with run finalization.
• run (trace_archive.rs): builds the .tar.gz with pure-Rust tar + flate2 (miniz_oxide backend) — no temp dir, no subprocess.

Trace archiving is best-effort: a failure logs but never fails an otherwise-successful check run.

Archive layout

01-requirement-slug/
  check-title-slug.attempt-1.jsonl
  check-title-slug.attempt-2.jsonl   # a retry
02-another-requirement/
  other-check.attempt-1.jsonl

One directory per requirement, one file per check execution, numbered by retry.

Testing

• cargo make clippy — clean (incl. --all-targets).
• cargo make test — 106 tests pass, including 4 new: TraceRecorder coalescing + NDJSON framing + tool-result truncation, and a tar.gz round-trip asserting exact archive paths (incl. same-title collision disambiguation).
• Verified --trace-archive appears in multi check --help.
• Not exercised end-to-end here (needs an Anthropic key + a CHECKS.md): a live agent run producing a real archive.

Stacked on #149.

🤖 Generated with Claude Code

[RobbieMcKinstry]

• feat(check): retire the thinking-disabled workaround — effort buys extended thinking #153
• fix(check): stop check agents from getting lost outside their sandbox #152
• fix(check): stop multi check from hanging on exit after a run completes #151
• feat(check): opt-in session-trace archive for multi check #150 👈 (View in Graphite)
• fix(check): stop the presenter and CLAUDE.md loading from corrupting output #149
• trunk

This stack of pull requests is managed by Graphite. Learn more about stacking.