refactor(core): share subagent job execution with recovery

[kitlangton]

Why

Live subagent execution and restart recovery each build their own job and select its final answer differently: the tool reads the latest 20 messages, while recovery scans the current model context. Maintaining both definitions makes restart behavior diverge from the normal path.

Execution also retains a UserInterruptedError variant that production never constructs. User interruption already travels through fiber interruption and the coordinator's explicit reason.

What Changes

• Both entry points call SubagentJob.start, sharing child resume, job metadata, and result selection.
• Recovery follows the existing live rule: select the newest completed, non-error assistant message among the latest 20 messages, concatenate only its text parts, and use the shared no-text fallback when needed.
• Recovered jobs keep their persisted job and notification IDs. Resume failures remain job errors; notification admission and acknowledgment stay with their existing owners.
• Remove the unused error class, runner error-union member, and classification branches. Actual user/shutdown interruption behavior is unchanged.

flowchart LR
  Tool[Subagent tool] --> Start[SubagentJob.start]
  Restart[Restart recovery] --> Start
  Start --> Job[Job.start]
  Job --> Resume[Session.resume]
  Resume --> Result[Shared result selection]

Loading

Scope

This consolidates subagent job execution and removes one dead interruption path. It does not add leases, change coordinator ownership, or redesign recovery accounting. Child-claim terminal events (#46943) and the claim-column rename (#46945) are separate changes. Land the rename last and reconcile its overlapping child-claim query with #46943.

Verification

cd packages/core
bun run test test/session-execution.test.ts test/tool-subagent.test.ts test/session-run-coordinator.test.ts test/job.test.ts test/session-error.test.ts
bun typecheck

96 tests passed; Core typecheck passed, using the repository-pinned Bun 1.4.0. Result-selection tests exercise live and recovered jobs against real persisted messages, including failed/incomplete candidates, empty output, and a result outside the 20-message window. Existing tests cover interruption, recovery accounting, job lifecycle, and notification deduplication.

The unmodified pre-push typecheck hook also passed: 33 successful tasks, 27 cached. No live provider/server end-to-end run was performed.