Add stage checkpoints and resume to SourceHuntRunner

[jorge-garcia-le]

Summary

Adds per-stage checkpoints and a resume capability to a single SourceHuntRunner run. Today, if a run crashes partway through, everything is lost — the hunt runs entirely in memory and findings.json/SARIF/report.md are written only once at the end. This change writes a lossless findings snapshot at each natural pipeline phase boundary so a crashed run can resume from the last completed phase.

Two goals:

• (a) Eval transparency — a physical, json.load-simple artifact per phase showing how findings evolve across hunt → verify → exploit.
• (b) Resume — re-run a crashed session from the last completed phase instead of from scratch, carrying budget forward.

Scope is deliberately coarse (three phase-level checkpoints, not the ~10 internal sub-stages). Mid-hunt/per-file resume, the proof flow, and changes to existing artifacts are out of scope.

What it does

Three checkpoints under <output_dir>/<session_id>/checkpoints/:

File	Written	Payload
hunt.json	after per-file hunt + static-merge + subsystem hunt	findings
verify.json	after verify + stability	findings + verified + rejected
exploit.json	after exploit	findings + verified + rejected + exploited

Each file is a superset of findings.json. Serialization is lossless (dataclasses.asdict, not the pool's lossy discovery-time subset); reload uses the tolerant __dataclass_fields__ filter.

Resume via a new resume_session parameter (accepts a bare sh-<uuid> name or a path to a session dir):

• Highest checkpoint present wins (exploit > verify > hunt).
• Re-runs preprocess (rebuilds repo checkout / callgraph / sandbox that verify+exploit need); skips rank — its priority fields are only consumed by the hunt phase.
• Guards each completed phase and its enrichment/side-effect sub-stages so only work after the last checkpoint executes.
• Report always re-runs (cheap; regenerates findings.json/SARIF/report.md).
• Deliberate exit: a session dir with no checkpoint raises rather than silently starting over.

Budget carries forward: the prior run's settled spend is summed from the existing spend-ledger.jsonl and seeded into a new SpendLedger via a new initial_spent_usd kwarg, so a resumed run honors the original dollar cap.

Existing artifacts (findings.json, findings_pool.jsonl, etc.) are untouched — only new checkpoint files are added.

Changes

• New clearwing/sourcehunt/checkpoints.py — standalone checkpoint read/write (atomic mkstemp + os.replace), latest_checkpoint, sum_prior_spend, resolve_session_dir.
• clearwing/llm/budget.py — new initial_spent_usd kwarg on SpendLedger (validated finite/≥0), recorded as carried_forward_usd in the run_started event.
• clearwing/sourcehunt/runner.py — resume_session param; adopts the resumed session id before instrumentation; seeds the ledger on resume; preflight skips models for skipped phases; arun writes the three checkpoints and guards completed phases on resume.
• tests/test_sourcehunt_checkpoints.py — 27 tests covering lossless round-trip, priority ordering, spend reconstruction, budget seeding, deliberate exit, and end-to-end resume (hunt+verify skipped, report re-runs, budget carried).

Eval passthrough

No call-site change needed in agent_eval.py — resume_session flows through the existing **extra/--runner-arg resume_session=<name> passthrough.

Verification

• 27/27 new tests pass; 35/35 existing budget + runner regression tests pass.
• ruff check clean on all touched files.