fall back to a local reviewer's served model so an unpinned mtplx review isn't blocked

[atomantic]

Summary

• A review loop configured with mtplx and no mtplxModel scalar failed every pass with No model configured for mtplx reviewer — which the agent's local-review script maps to no-verdict, i.e. review-blocked — while the MTPLX daemon was up and serving exactly one model.
• runToolFreeLocalCompletion now resolves an unpinned model from the backend's own /v1/models listing (via the shared probeOpenAiModels) and uses it when the backend serves exactly one — the single-model-daemon case (MTPLX, llama.cpp, vLLM, LM Studio with one model loaded).
• Ambiguity is never guessed: several models (an ordinary Ollama install lists every one it has), none, or an unreadable listing still error — now naming which of those it was, instead of a bare config message.
• The no-model failure carries code: 'NO_MODEL', so a config gap maps to 400 at POST /api/code-review/local and to RECHECK_NO_MODEL in the challenge re-check, rather than the 502 bucket meant for a reviewer that ran and failed.
• Results (and the re-check's resolution note) report the model the pass actually ran with, not the absent argument.
• Drive-by fix: the challenge re-check read its model from an ollama-or-lmstudio ternary, so any third local backend got LM Studio's model id. It now reads <backend>Model, matching the route.

Test plan

• cd server && npm test — full suite green (1908 files).
• New coverage in server/services/codeReview.test.js: single served model is used (and the MTPLX /v1 root isn't doubled in the probe URL), several models refuse to guess with no review request sent, an unreachable backend says so, and an up-but-empty backend still asks for a pin.
• New coverage in server/routes/codeReview.test.js (400 on NO_MODEL) and server/services/cosTaskStore.test.js (per-backend scalar, unpinned delegation, RECHECK_NO_MODEL passthrough).
• Verified live against the local MTPLX daemon: an unpinned mtplx review resolved the served model and returned real findings on a seeded typo diff.