fix: consolidate open correctness fixes

[dylanneve1]

Summary

Consolidates the valid/current correctness fixes from the remaining stale open PRs onto current main after the Agy/Antigravity removal. I reviewed the old PR intent and ported the pieces that still apply instead of taking the branches wholesale; stale backend-specific changes and broad logging/bootstrap churn were intentionally left out.

Key areas covered:

• Fix built-in OpenAI Agents lifecycle/discovery behavior on Windows by using PowerShell for the built-in Bash tool and skipping bash trigger probes when no usable bash exists.
• Honor retry delays and preserve valid chat_id=0 routing.
• Make Claude flow violations retry rather than silently drop turns.
• Harden cron/session/trigger/history storage paths against null and migration edge cases.
• Preserve dream last_run during running/failure states.
• Clean up pulse disable registration state.
• Tighten Teams, Telegram, terminal, CLI doctor, and terminal chat correctness issues.
• Increase Windows timeout tolerance for the flaky backend/model handler tests.

Supersedes

Intended to supersede/replace #232, #220, #212, #205, #200, #168, #162, #152, and the still-relevant low-risk parts of #90.

Validation

• npm run typecheck
• npm run lint (passes with existing warnings)
• Prettier check over touched files
• npx vitest run --exclude src/__tests__/package.functional.test.ts --reporter=dot
• Targeted tests for OpenAI Agents discovery, triggers, Teams, Claude watchdog, backend registry, Codex handler, and related changed areas

The full package install smoke test was not rerun successfully on this Windows machine because local disk space hit ENOSPC; the broad suite excluding that package smoke passed.