test(taskctl): improve pipeline.test.ts to cover processAdversarialVerdicts state machine

[randomm]

Context

Phase 3b (PR on feature/taskctl-pipeline) implemented the developer→adversarial→commit inner loop in packages/opencode/src/tasks/pulse.ts. The tests in packages/opencode/test/tasks/pipeline.test.ts currently only verify that adversarial verdict data is correctly stored in the task store — they do not test the actual state machine logic.

What's missing

The following pulse.ts functions have no meaningful test coverage:

• processAdversarialVerdicts() — APPROVED path should result in task status closed + PM notification fired; ISSUES_FOUND should increment pipeline.attempt and respawn developer; 3rd failure should set status failed and escalate
• commitTask() — happy path closes task; failure path calls escalateCommitFailure
• respawnDeveloper() — new session created in same worktree, attempt counter incremented, adversarial feedback passed in prompt
• escalateToPM() — task status set to failed, BackgroundTaskEvent published
• spawnAdversarial() — spawned when pipeline.stage === 'reviewing', double-spawn prevention via adversarial-running stage

Why it's hard

These functions call Session.createNext(), SessionPrompt.prompt(), and Worktree.create/remove — real infrastructure that's difficult to test without a running server. The tests were scoped to Store-level data assertions as a pragmatic choice.

Suggested approach

Mock or stub Session and Worktree at the module boundary. Options:

1. Extract the state-transition logic into pure functions that take/return plain objects (easier to test)
2. Use Bun's module mock capabilities to stub Session.createNext and SessionPrompt.prompt
3. Create a lightweight integration test harness that sets up a minimal in-process session

Acceptance criteria

• processAdversarialVerdicts: APPROVED → task closed test
• processAdversarialVerdicts: ISSUES_FOUND → attempt incremented, developer respawned test
• processAdversarialVerdicts: 3rd failure → task failed + escalation test
• Double-processing prevention: verdict cleared before action (regression test)
• bun test passes, bun run typecheck passes