fix(delegation): settle stopped subagents reliably

[wen2zhou]

Problem

A user-reported regression caused stopped subagents to lose output that had already been produced, affected delegated ask-user cards, and allowed notebook host.stopChild to block indefinitely when provider, artifact, or lease cleanup did not settle.

Proposed change

• Separate the cancellation action fence from the runtime observation lane. Stop blocks new continuation, permission, question-admission, and capability writes without discarding same-Attempt runtime observations.
• Persist same-Attempt late message and tool evidence through transport shutdown, with event-id idempotency, so the evidence remains visible after reopening the Session.
• Own transcript projection through per-Turn lanes keyed by runtime segment and prompt Message. Live observation and terminal settlement reuse one Message identity without duplicating output or attaching evidence to another Branch.
• Keep terminal lifecycle authoritative: late evidence can extend the transcript but cannot revive or rewrite a completed, cancelled, or errored Attempt.
• Keep a completed subagent's pending ask-user card answerable. Stop atomically cancels pending questions while transitioning only a still-running Attempt.
• Acknowledge durable Stop independently from detached provider, artifact, backend-claim, and capacity cleanup, while preserving observable retry and exactly-once release behavior.
• Keep cleanup reporting, detached drains, and retryable exactly-once releases behind the delegated-cleanup deep Module instead of expanding the delegated-work composer.
• Keep durable/live transcript reconciliation behind the workspace-subagent-runtime-transcript deep Module, leaving the runtime presentation hook focused on subscription and lifecycle fencing.
• Preserve the user-facing Cancelled label and warning-color status semantics.

Scope

• Delegated ACP cancellation, observation, and cleanup lifecycle.
• Per-Turn delegated transcript persistence and Session reload behavior.
• Durable delegated-work Stop/question finalization and notebook RPC coverage.
• Renderer subagent summary/detail terminal-state projection.
• No documentation changes.

Acceptance

• Cancelled, failed, and awaiting-user subagents do not retain stale Thinking state.
• Stop fences actions without dropping same-Attempt late message or tool evidence.
• Persisted late evidence remains visible after reopening and cannot revive terminal lifecycle state.
• Multi-Turn Attempts retain one transcript Message identity per Turn lane without duplicate output or cross-Branch evidence.
• A completed Attempt may retain an answerable pending ask-user card; Stop cancels that card atomically without rewriting the completed Attempt.
• host.stopChild returns without waiting indefinitely for external cleanup.
• Independent ownership releases remain observable, retryable, and exactly once.
• The terminal UI label remains Cancelled.

Review focus

• Separation of the action fence from the observation lane.
• Per-Turn transcript lane identity, event-id idempotency, and persistence provenance.
• Atomic question cancellation versus Attempt lifecycle transitions.
• Separation of durable Stop acknowledgement from external cleanup drains.
• Renderer preservation of durable output while clearing only transient runtime presentation.
• The cleanup and renderer transcript Module seams, including their reduced composer/presentation surfaces and module-impact ownership.

Related issue

User-reported regression; no issue number.

Verification

All checks below ran after the last material edit:

• Main Stop/persistence/architecture focused set: 5 files passed, 77 tests passed.
• Delegation, ACP, transcript, and notebook focused set: 4 files passed, 152 tests passed.
• session_persistence Module suite: 6 files passed, 265 tests passed.
• Renderer architecture, HomePage, module-impact, and release surfaces: 7 files passed, 197 tests passed.
• Module-impact validator/shadow rerun: 2 files passed, 22 tests passed.
• npm run typecheck:node
• npm run typecheck:web
• ESLint on every modified/new TypeScript and TSX file.
• Prettier check on every modified/new file.
• Working and cached git diff --check with no whitespace errors.

The known renderer callable-surface count mismatch, stale local Prisma client failures, and Windows E2E failure were independently identified as unrelated to this PR and were not changed here.