Fixes #1788

Summary

• seed runner workspaces from a filtered external snapshot, avoiding recursive workspace/artifact mounts
• treat validated changed-files and patch evidence as actionable generic execution output without fabricating typed or package artifacts
• emit repository-relative canonical patch headers so the host workspace patch applies safely
• add real CLI Playground E2E coverage for workspace application, publication, refs, exclusions, and privacy
• centralize cleanup of the private runtime source root and the runner workspace seed snapshot into one idempotent coordinator that every normal, failure, and finally path routes through exactly once; SIGINT/SIGTERM/SIGHUP await the same coordinator before exiting with the conventional 128+signum code, with a bounded synchronous best-effort fallback on process exit

Cleanup coordinator (final blocker)

• Source relationship: replaces the four previously independent cleanup sites in .github/scripts/run-agent-task/execute-native-agent-task.mjs (mid-run awaits, an inline catch-path duplicate, a partial exit hook covering only the runtime source root, and signal handlers that cleaned only one resource and exited 128 for every signal).
• Change kind: refactor of existing cleanup behavior plus a signal-handling correctness fix; no new runtime capability. The runtime change is not broader than the evidence: the same two temp roots are removed, only ownership moves to one coordinator and signal exits gain conventional codes.
• Verification capability: tests/execute-native-agent-task-interruption.test.mjs deterministically starts the executor in a child process, pauses it immediately after seed snapshot creation via a narrowly scoped NODE_ENV=test marker-file hook (inert in production), delivers SIGTERM/SIGINT/SIGHUP, and asserts the temp seed snapshot directory is removed, the exit code is 143/130/129 respectively, and no runner workspace content survives; a normal run in the same harness proves routine cleanup still passes. tests/runtime-sources-materialization.test.ts pins the coordinator shape: one coordinator definition, signal handlers routed through it, a single top-level finally, and no duplicate independent cleanup logic. Error reporting/redaction is preserved (stderr failure messages are now redacted while the private root is still known).

How to test

1. Run npm run build.
2. Run npm run test:agent-task-contracts.
3. Run npm run test:native-agent-task-playground-e2e.
4. Run npm run test:native-agent-task-interruption and confirm it reports native agent task interruption ok after exercising SIGTERM, SIGINT, and SIGHUP plus a normal completion.
5. Confirm the E2E reports native agent task Playground e2e ok; it verifies host README.md application, publication projection, canonical refs, empty typed artifacts, prepared workspace mounts, snapshot exclusions, and redaction.

Backwards compatibility

• Agent task results with no validated file changes remain incomplete/no-op as before. Valid changed-file and patch evidence now allows an otherwise empty semantic workload to succeed.
• Signal-interrupted executor runs now exit with the conventional 128+signum code (130/143/129) instead of a uniform 128. Only consumers matching the exact previous exit code 128 for signal deaths would observe the difference; the repository's own workflow treats any non-zero exit as failure.

AI assistance

• AI assistance: Yes
• Tool(s): openai/gpt-5.6-terra and OpenCode; Claude Opus 4.5 via OpenCode for the cleanup coordinator, signal handling, and interruption test
• Used for: Implemented and verified the workspace snapshot, normalization, Playground E2E changes, and the centralized signal-safe cleanup with Chris-directed requirements.