Persist coding task execution and completion timing

[witbrock]

Merge decision: ready — the canonical task can retain actual coding-attempt start/end observations and accepted completion, with idempotent read-back through the existing controllers. The operator-owned VS Code integration is supplied as a reviewed patch; neither host activation nor public deployment is claimed.

User outcome

Coding-task effort can be measured from execution rather than queue acceptance or generic task updates. Task reads and completion receipts distinguish execution end, successful completion acceptance and elapsed wall-clock duration (including waits).

Native assignment: #V#task_agent_8b706f7c6efb91f3241fb40bc0de1bfb.

Material changes

• Store observations under canonical task metadata with existing attempt IDs, compare-and-set updates, immutable result receipts, and retained completion history across reopening. Unknown historical/exit times stay unknown.
• Record DGX execution after subprocess creation, preserve exit evidence on recovery, and include canonical timing in result messages while preserving supervisor-repair ordering.
• Supply the bounded patch for the existing operator-owned Mac bridge's begin/resume/finish actions. Its task/input checks and existing consumer remain authoritative. No new consumer or migration is introduced.
• Expose timing through existing task detail/list reads and HTTP detail. Cost is explicitly unknown; subscription token usage is not an attributable per-task invoice. A separate native cost-backlog creation request is retained for the controller, which alone owns native effects in this run.

Evidence

• 188 passed across task execution timing, DGX worker/retry/supervision, canonical task service and run archive tests.
• Tests include a real local fixture subprocess through DGX launch/result, canonical HTTP detail retrieval, retry/resume, duplicate delivery, lost acceptance acknowledgement, reopening, unknown end/cost, and changed-input withholding.
• Isolated Mac replay applied the patch to inspected operator source SHA-256 a665516dfe918b6e77046867721ad471c787535bf5e8a92d7aa677c57f19df5b, executing only its function definitions against fixture state and canonical services on a mock database. No host bootstrap, real model, live task or messaging helper ran.
• Python 3.11 grammar compatibility passed for 1,580 files; focused Ruff checks passed.
• An additional task-detail external-resource test returns 404 on both this candidate and unchanged baseline 09b913954; its mocked task is absent from the deployment lookup. It is unrelated and is not included in the 188 passing affected tests.

Ship boundary

Minimum repository evidence is the persisted lifecycle/read-back and both adapter code-path replays above. False completion times, erased history, repeated completion of a reopened task or broken existing execution would block this change. The operator must separately apply the reviewed bridge patch and select a coherent worker/backend release under existing locks, retaining rollback, before claiming activation. The patch replay is optional in CI because its source is operator-owned; the remaining timing tests run without it.

No public deployment is requested. Billing integration, historical timestamp invention, federation/import changes and additional task consumers are outside this change.