`keep_workspaces = false` never reaps task worktrees

[mrthankyou]

Summary

With keep_workspaces = false, a fully successful run still keeps every task worktree. Smithers logs warn keeping worktree with unsaved work once per task, because each worktree contains files Ultrafuzz itself wrote there. ultrafuzz clean <run> then deletes the run directory but leaves the Git worktree registrations, the per-task branches and the Smithers run record. Relaunching with the same run ID fails at admission after ~7 minutes of preparation.

Evidence (main 6ac8c2bf, 2026-09-29)

Reproduced in two smoke runs against Damn Vulnerable DeFi v4.1.0 (Claude, 8/8 nodes succeeded). The second ran on 6ac8c2bf plus an unrelated agent-failure-reporting change.

Worktrees are never reaped

• All 8 task worktrees remain under <run>/workspaces/, and the engine log contains warn keeping worktree with unsaved work 8 times.
• Every worktree's git status shows only Ultrafuzz-owned paths:

  ?? .ultrafuzz/
  ?? artifacts/
  

  Strategy nodes additionally show ?? test/foundry/ and, in one case, ?? foundry.lock.
• Worktrees were 19 MB and 104 MB in the two runs. Real targets with out/, cache/ or node_modules/ will be much larger, and the registrations accumulate with every run.

clean leaves Git and engine state behind

• ultrafuzz clean <run> --yes removes the run directory and refs/ultrafuzz/runs/<run>/source, and nothing else (packages/runtime/src/clean.ts).
• Afterwards git worktree list still has 8 prunable entries, and the 8 ultrafuzz/<run>/<node> branches remain.
• The project-root smithers.db still has the run in _smithers_runs (ultrafuzz-<run> | finished).
• ultrafuzz run --run-id <run> then rebuilds the plan and the ~1 GB execution snapshot, and only at submission fails with:

  WORKFLOW_SUBMISSION_FAILED: workflow runner command failed (exit 1)
  DETACHED_ADMISSION_FAILED: Detached engine exited before admission (exit 4)
  code: RUN_EXISTS
  message: "Run already exists: ultrafuzz-<run>"
  

  The failed launch leaves a new partial run directory behind.
• Not yet tested: whether the leftover branches and registrations also break git worktree add once the Smithers record is gone. Admission fails before any worktree is created.

Proposal

• After verification publishes a task's artifacts, remove Ultrafuzz-owned paths (.ultrafuzz/, artifacts/) from the worktree. Alternatively, reap with Ultrafuzz's own cleanup instead of relying on Smithers' unsaved-work check. Keep a worktree only when it holds changes outside Ultrafuzz-owned paths.
• clean should also remove the run's worktree registrations, ultrafuzz/<run>/* branches and Smithers run record.
• Independently, run should reject a run ID that Smithers already knows before it builds the snapshot.

Acceptance

• A successful run with keep_workspaces = false leaves no task worktrees and no stale git worktree list entries.
• After clean <run>, run --run-id <run> launches successfully.

Related: #122 (artifact preservation before reaping, closed), #223 (worktree cleanup permissions, closed), #988 (clean leaking snapshots), #1148 / #1166 (source pinning).

[aviggiano]

Rescoping this. #1259 is closed: writing .gitignore files into worktrees made them look clean, so keep_workspaces = false could delete worktrees the post-run gate still reads, plus the only copy of a rejected output. We've also seen many failed runs tied to workspaces and declared outputs, so I'd rather fix the lifecycle once than add another layer.

New scope: Ultrafuzz deletes a task worktree itself once its outputs are published and the post-run checks no longer need it, instead of relying on git status being clean. It never deletes the only copy of a rejected output. The clean part is #1257, and the run-ID part was fixed by #1264, so both are out of scope here.

Done when:

• a smoke run with keep_workspaces = false leaves no worktree whose only changes are Ultrafuzz-owned;
• its report stays verified;
• a failure_policy: continue task with a rejected output keeps that output;
• the strategy-worktree policy (test/foundry/, foundry.lock) is decided and documented.

[aviggiano]

#1282 implements this scope. After a run ends, Ultrafuzz deletes each task worktree whose outputs are published. A worktree that holds a rejected output is kept. The engine no longer reaps worktrees. The CLI e2e campaign now asserts that workspaces/ is empty after status and that the report stays verified. test/foundry/ and foundry.lock are deleted with the worktree (documented in docs/reference/configuration.md).

@mrthankyou, your work in #1259 is credited as co-author on #1282.