fix: persist ACP model and reasoning effort across turns and resume

[superbiche]

This refreshes @gilbertl's #337 onto current main (v1.7.0), preserving the original commits and authorship while resolving the conflicts. The scope is intentionally narrower than the earlier revision: model and reasoning-effort persistence only; agent-mode behavior remains identical to upstream main.

What changes:

• model and reasoning_effort configuration updates are persisted with thread/settings/update and acknowledged only after Codex accepts them. A setting therefore remains effective for subsequent turns and process reconnects until changed.
• Resume, load, and fork no longer invent modelProvider: "openai" when no provider is configured. Omitting it lets the Codex thread retain its stored model and reasoning effort.
• Regression coverage exercises persistence across an adapter restart and session load.

Validation on the final candidate (0275316):

• npm run typecheck: pass
• npm test: 486 passed, 27 skipped
• npm run bundle:all: all six platform bundles pass
• npm run build: pass
• Live ChatGPT-auth reconnect oracle: selected effort changed high → low; after two independent adapter restarts, both resume and load reported low
• Independent deep review: three rounds, final PASS; agent-mode persistence was removed during review instead of adding concurrency complexity unrelated to this fix

Known boundary: concurrent ACP callers can still race a setting update against a prompt already in flight. The deployed acpx flow is sequential and waits for the setting acknowledgement before sending the next prompt.

Transport-level effective-config receipts need an authoritative app-server receipt, and per-invocation effort for one-shot acpx exec belongs in acpx; neither is bundled into this adapter fix.

Fixes #336.
Fixes #343.

---

_{@superbiche · user · drafted with Codex (GPT-5).}

[superbiche]

This conflicted branch has been superseded by #457, rebased onto current main / v1.8.0 with the persistence, prompt-serialization, resume/load/fork, and regression-test updates intact.

I left this PR open rather than rewriting its published history; maintainer review should continue on #457.

---

_{@superbiche · user · drafted with Codex (GPT-5).}