refactor(fb15c): split SubagentRunner into modules

[alexdim]

Summary

Split the 1033-line SubagentRunner.ts into focused modules under src/core/task/tools/subagent/:

• SubagentRunTypes.ts — run status/result/progress/tool-call types
• SubagentRunHelpers.ts — pure functions (tool-call parsing, result blocks, best-effort result)
• SubagentRunState.ts — phase/liveness/heartbeat tracking
• SubagentRunProgress.ts — progress queue/drain + recorder flush
  SubagentRunner.ts becomes the orchestration adapter.

Why

FIX-BACKLOG FB-15c — single responsibility: runner keeps orchestration; state, progress, types, and pure helpers are separate units.

Review follow-ups

1. Flush blocking terminal completion

Reviewer flagged await this.runProgress.flush() in the finally block: it changed recorder flushing from fire-and-forget to part of the awaited run lifecycle, so a slow or stuck filesystem append could delay terminal completion.

Fix: restored the pre-split fire-and-forget behavior — void this.runProgress.flush() instead of await. Errors are still handled internally by SubagentRunProgress.flush() (try/catch → reportRecorderFailure).

Regression test: does not block terminal completion on a slow recorder flush — stubs recorder.flush() to hang forever; asserts run() still resolves with COMPLETED and that flush() was called.

2. Terminal phase stuck on "cancelling"

Reviewer flagged that recordTerminal was missing the enterPhase call after extraction. On timeout/parent-cancel, the terminal progress reported status: cancelled while phase remained cancelling.

Fix: restored this.enterPhase(this.phaseForStatus(result.status), "subagent run settled", { status: result.status }) before runProgress.recordTerminal() in SubagentRunner.recordTerminal().

Regression test: reports cancelled phase on parent cancel, not cancelling — triggers parent abort, collects progress updates, asserts every cancelled-status update has phase: "cancelled".

Notes / Caveats

• Rebased onto current master.
• Pre-existing subagent drain-timeout test failure is unrelated and untouched.
• Pre-existing AgentConfigLoader EPERM on TCC-protected ~/Documents/Dirac is environment-specific (machine has no Full Disk Access); not from this PR.

Verification

• tsc -p tsconfig.unit-test.json --noEmit — only pre-existing src/shared/storage/types.ts baseline error.
• biome check on touched files — 1 pre-existing warning (not from these changes).
• SubagentRunner.test.ts — new regression tests passing (2/2 targeted).
• SubagentRunRecorder.test.ts — 4/4 passing.
• SubagentBuilder.test.ts — 5/5 passing.

Closes FB-15c (FIX-BACKLOG).

[dirac-run]

Requested follow-up: await this.runProgress.flush() changes recorder flushing from fire-and-forget to part of the awaited run lifecycle. A slow or stuck filesystem append can now delay terminal completion, despite the PR being described as behavior-preserving. Please either preserve the previous non-blocking behavior or explicitly document and test the new completion contract, ideally with a bounded wait.

[dirac-run]

Follow-up needed:

The recorder flush issue is fixed, but the extraction dropped the terminal phase transition from SubagentRunner.recordTerminal().

When timeout or parent cancellation wins Promise.race, the result is built directly without passing through instrumentedOnProgress. Terminal progress therefore reports status: cancelled while phase remains cancelling.

Please restore enterPhase(phaseForStatus(result.status), ...) before recording terminal state and add a regression test for the timeout/parent-cancellation path asserting a final cancelled phase.

[dirac-run]

LGTM.