feat(codex): support ultra without resetting sessions [AGENT: miro]

[utiberious]

Ports upstream PR chenhg5#1519 reasoning/model support onto the production fork and fixes /reasoning session loss.

Changes:

• supports none, minimal, low, medium, high, xhigh, max, and ultra
• preserves live and idle resumable Codex conversations across /reasoning changes
• applies the new effort on exec resume and app-server turn/resume
• serializes live effort updates with concurrent session cleanup
• fixes multi-workspace reasoning card state
• adopts upstream dynamic chat-model discovery dependency (fix(codex): surface gpt-5.x in /model chooser (drop stale allowlist) chenhg5/cc-connect#1546)
• adds unit, race, and CUJ coverage for chat -> /reasoning ultra -> chat

Compatibility caveat:

• reasoning levels are currently advertised globally; support depends on the selected model, and choosing an unsupported level can make the next turn fail. This is documented in config.example.toml.

Verification:

• go test ./agent/codex ./core -count=1
• go test ./core -run TestCUJ -count=1
• go test -race ./core -count=1
• concurrent cleanup regression repeated under race detector
• make build-noweb
• real Codex CLI low -> ultra resume retained the same thread ID and recalled its marker
• independent post-fix review found no remaining blocking issue

No deployment or restart is included.