fix(providers): probe ollama thinking capability before sending reasoning_effort on custom endpoints

[tomRumi]

Problem

Ollama's /v1/chat/completions rejects top-level reasoning_effort with HTTP 400 "<model>" does not support thinking whenever the target model lacks the thinking capability (verified live: granite4:3b, ministral-3:8b; their /api/show capabilities are [completion, tools] / [completion, vision, tools]).

CustomProfile forwarded any configured effort verbatim, so a global agent.reasoning_effort: high broke every non-thinking local model: three retries, failed turn, and no hint in the UI — the error only surfaced in errors.log. The workaround (agent.reasoning_overrides.<model>: disabled) must be re-done by hand for every non-thinking model the user ever pulls.

The main conversation path and the auxiliary path (session-title generation, which resolves through the same profile) both hit it:

ERROR agent.conversation_loop: API call failed after 3 retries.
HTTP 400: "granite4:3b" does not support thinking | provider=custom:ollama-(local) model=granite4:3b

Fix

CustomProfile reuses the existing hermes_cli.models_local.ollama_model_supports_thinking() probe (native /api/show capabilities — the same authority OllamaCloudProfile and ReasoningParamsMixin._ollama_supports_thinking_cached() already rely on) for Ollama-shaped custom endpoints before emitting an effort:

probe result	behaviour
thinking in capabilities	effort forwarded verbatim (unchanged for qwen3, deepseek-r1, …)
probed OK, no thinking	reasoning_effort omitted — request cannot 400
probe unreachable / raises	fail open — effort sent (worst case is the pre-fix behaviour; a transient probe failure must not silently drop the user's reasoning config)
non-Ollama custom endpoint (GLM/ARK, Groq, vLLM, llama.cpp)	no probe, no latency, no behaviour change (_looks_like_ollama_endpoint gate; #57601 passthrough contract intact)

Definitive results are cached per (model, base_url) on the profile singleton — one /api/show per model per process, not per request build; failed probes retry after a 60 s TTL (mirrors the _cached_probe policy in reasoning_params.py). The explicit-disable path (reasoning_effort: none + think: false, #14820/#25758) is untouched and stays probe-free.

Testing

• 9 new invariant tests in tests/plugins/model_providers/test_custom_profile.py: omit-on-non-thinking, verbatim-on-thinking, fail-open on None-returning and raising probes, definitive caching (+ distinct key per model), no probe on non-Ollama endpoints, disable path still probe-free.
• Full file: 49 passed (tests/plugins/model_providers/test_custom_profile.py + test_ollama_cloud_profile.py), existing 40 assertions unchanged and green.
• E2E against a live ollama 0.12.10 on this machine: with agent.reasoning_effort: high and no overrides, granite4:3b now answers a one-shot hermes chat -q turn (previously: 3x retry, then HTTP 400 failure).

Notes for reviewers

• Ollama's native /api/chat happily ignores thinking params; this only concerns the OpenAI-compat /v1 route Hermes uses for custom providers.
• An alternative "catch the 400 and retry without the field" recovery exists in spirit (_ephemeral_reasoning_off) but only covers disable-rejected routes; probing before sending is cheaper than a failed turn + retry cycle, and matches how Ollama Cloud is already handled.

[alt-glitch]

This was generated by AI during triage.

Related: this is the fourth open fix for #59660 alongside #95854, #81574 and #86197 (plus closed #63315). All gate reasoning_effort on the Ollama /api/show thinking capability but differ in where the probe lives (this PR reuses hermes_cli.models_local.ollama_model_supports_thinking). Flagging the cluster so a maintainer can pick one mechanism.