fix: resolve ollama thinking support per model before sending reasoning_effort

[atomantic]

Summary

A local reviewer pinned to ollama[<non-thinking model>]~effort=low sent reasoning_effort on its first request. Ollama translates that field into its own thinking parameter and hard-rejects the whole request with a 400 for a model that does not advertise thinking support.

PR #5960 added a reactive recovery (retry without the field), but the "this model has no thinking" memo lives in a process-local Map — and every reviewer invocation from a claim/PR run is its own short-lived node process. So the cache never survived to the next call, and every review re-uploaded the entire diff twice: once for a request that could never succeed.

Ollama's /api/show reports the capability set per model, and PortOS already reads it (ollamaManager.getModelCapabilities, cached). This resolves it before the request and omits reasoning_effort for a model lacking thinking, so the doomed request is never sent. The 400-retry stays as the fail-safe for a backend with no such probe (LM Studio, MTPLX) and for a probe that cannot answer.

Sentinel handling follows the repo's absent-vs-empty rule: null (probe failed) and [] (daemon answered reporting no capabilities) both mean unknown, not unsupported, so both fall through to the request rather than silently stripping a level a reasoning model does accept.

The probe is skipped entirely when no effort is pinned — nothing to drop, so a capability round-trip buys nothing. That also stops effortUnsupported: true being reported when the caller never asked for a level.

The runLocalCodeReview doc comment now states the omission is decided per model, not per backend, matching the behavior (this was called out in the issue as the contract the implementation had drifted from).

Test plan

• server/services/codeReview.test.js — 6 new cases pinning the body shape both ways and the fall-through paths:
  • reasoning_effort absent on the first request when /api/show omits thinking (fails on main, passes here — one fetch, not two)
  • reasoning_effort present when /api/show reports thinking
  • [] capabilities treated as unknown → level still sent
  • probe unable to answer → reactive 400-retry still recovers
  • no effort pinned → no probe, no effortUnsupported flag
  • non-ollama backend → not probed, level still sent
• Full server suite green: 1896 files / 38311 tests passed, 1 file + 14 tests skipped.
• Verified end-to-end against the live ollama daemon with the exact model that reproduced the bug: previously ok:false with 400 ... does not support thinking; now ok:true with findings and no retry warning, confirming the first request was already correct.

Closes #6050