Prevent unsafe launches and recover native sessions correctly

[prithvi2206]

Problem or objective

• Address the correctness and operational findings from the review of Put expiring AI subscription quota toward useful work #1 before unattended dogfooding.
• Project billing settings, disabled activity hooks, and configuration changes during probes could invalidate launch checks. Codex retries and usage-limit errors could also leave runs in the wrong state.

Solution

• Apply the same Claude subscription settings to authentication and execution in the skill's directory; reject an effective API key or non-subscription login.
• Verify native Codex hook enablement and trust, and track Claude parent and child activity separately.
• Require a terminal Codex result, retain retry diagnostics, and resume recognized usage-limit interruptions in the same session. Invalid or ambiguous results hold their slots.
• Reload configuration after asynchronous checks and before execution. Keep active weekly/manual sprints awake during quota waits, including before the first launch.
• Preserve native configuration directories in launchd, compare service ownership by the exact home argument, and avoid misleading backlog-recheck advice.

Design choices

• Reuse the existing sprint timing, SQLite store, and native runners. No new dependency or database migration.
• Keep hook approvals under user control and retain held slots for uncertain external work.
• Use temporary invocation settings without editing the user's billing configuration.

Validation

• All 68 tests and TypeScript checks pass on Node 22.22.0 and 24.4.1. Child-process tests cover both providers, continuation, configuration races, and malformed results.
• Native Claude authentication probes verified the billing overrides in an isolated project. Native Codex metadata probes verified hook flags, event names, canonical paths, and trust status without running model turns or approving hooks.
• PR Review Toolkit completed; its remaining initial-quota-wait finding is fixed and covered. Package dry-run excludes private state.
• Real prospect research, Starport preparation, native sleep/wake behavior, and a closing-time sprint remain release acceptance checks in docs/validation.md.