fix(codex): release active thread writers on dispose

[SDSLeon]

Summary

• recover authoritative in-progress Codex turn IDs during structured-session teardown
• interrupt every tracked or provider-reported active turn before unsubscribing
• preserve replacement sessions across ownership transfers and stale-ID interrupt retries

Root cause

Teardown relied on locally observed active turn IDs. If Poracode missed a concurrent or continuation turn, it could unsubscribe while that Codex turn kept running in a pooled app-server process. The provider thread retained its writer, so a later thread/resume failed with thread … already has an active writer.

Impact

Reopening a Codex GUI thread no longer intermittently fails because an unobserved turn survived teardown. Replacement sessions are protected from interruption if ownership changes while teardown is in flight.

Validation

• pnpm exec vitest run src/supervisor/agents/codex/codex.test.ts src/supervisor/agents/codex/appServerRpc.test.ts src/supervisor/agents/codex/serverPool.test.ts — 3 files, 161 tests passed
• pnpm run typecheck
• pnpm run lint
• pnpm exec oxfmt --check src/supervisor/agents/codex/acp.ts src/supervisor/agents/codex/codex.test.ts
• git diff --check