fix(sandbox): bind Docker controls and cleanup to run ownership

[ttbombadil]

Scope

Fix Docker lifecycle ownership on fresh origin/main d9a21de. Implements the approved internal-only decision. Wire schema, API version and operation_error operations remain unchanged.

Changes

• Pause/unpause admission remains boolean; WS awaits internal completion. State, deadlines, batching and clock markers change only after successful Docker completion.
• Controls capture run generation, abort identity, originating state and container. Stop/reuse/natural exit/actual Docker deadline invalidate late results. Current failure stops and releases via existing serial diagnostic and stopped status.
• Natural exit, deadline and Stop join runner-owned cleanup. Nonzero rm is not success: an allowlist-compatible successful daemon listing must confirm absence for --rm auto-removal.
• Failed cleanup retains its pool slot in quarantine. Retry uses the existing idle-maintenance interval; success restores the same runner. Shutdown retries stopped cleanup owners. Other pre-existing reset failure replacement behavior is retained.

Regression evidence

• Baseline RED: 11 of 15 runner cases failed; 3 WS cases failed. Review RED: auto-removal, natural-close control resurrection, failed-owner quarantine/recovery/shutdown. Additional RED uses actual DockerManager runtime deadline.
• New deterministic tests cover nonzero, spawn failures, timeouts, delayed unpause after Stop/reuse, stale failures, confirmation-only status, failed removal retry, quarantine and shutdown.
• Real Docker test inspects confirmed pause/unpause and checks same-runner reuse after natural --rm exit. Unique synthetic missing-container probe exercises real ProcessExecutor cleanup confirmation.
• Existing assertions unchanged. Two cleanup mocks now return the real successful ProcessExecution shape instead of undefined. Existing timing tests await control completion before unchanged PAUSED/Stop and time-freeze assertions; the freeze baseline is captured after confirmation, matching the approved requirement.

Verification

All local gates green: unit 2834 passed / 1 existing skip; integration toolchain 14 passed / 1 existing skip; Docker 28 passed; E2E 25 passed, no race events, zero retries; Typecheck, docs check, read-only lint, build and fresh coverage/Sonar quality gate passed. E2E uses unchanged assertions in temporary copies on port 3307 to avoid an unrelated service on port 3000. No no-verify, no real Provider calls, no SSOT/Tutor/Planner/Mastery/Evidence/prompt/content changes. Original untracked local documents and .env.local untouched. macOS Docker Desktop correctness tests are not representative capacity measurements.