fix(auto-run): unwedge experiment loop when provider rejects temperature

[ChenglongWang]

Summary

Production incident PRJ-0002 (2026-05-14): every azure/anthropic/claude-opus-4-7 call returned

litellm.BadRequestError: Azure_aiException - invalid_request_error: `temperature` is deprecated for this model.

Every auto-run round re-hit the same parameter error until _AUTO_MAX_ROUNDS (20) was exhausted, and the error string was surfaced to the user instead of any research output.

Logs that triggered this PR: ~/Shared/agent-service.log (19 identical _run_agent_loop:580 errors in 13 seconds) and ~/Shared/actions.jsonl (auto-run round 1..20 all firing on the same model with no recovery).

Three independent bugs collided. This PR fixes each in isolation.

Bug 1 — temperature attached to models that reject it

• New mira_engine/providers/model_compat.py centralises the rule (currently claude-opus-4-7).
• AzureOpenAIProvider._supports_temperature consults it on top of its existing gpt-5 / o1 / o3 / o4 blocklist.
• LiteLLMProvider.chat pops temperature from the request kwargs before model-specific overrides apply.
• AnthropicProvider._build_kwargs conditionally skips temperature on every code path (adaptive thinking, enabled thinking, plain).

Adding a model to the blocklist is now a one-line registry change.

Bug 2 — non-retryable raised exceptions burned the fallback chain

RoutedProviderManager.chat already classified non-retryable error responses via _should_retry_with_fallback, but the except Exception branch unconditionally walked to the next candidate. A raised BadRequestError therefore tried every candidate even though they all fail identically. Apply the same classifier on str(exc) and re-raise immediately on non-retryable errors.

Bug 3 — auto-run kept spinning on LLM provider errors

_evaluate_continuation only stopped on failure responses when strictHeuristics was on. In PRJ-0002 there was no automation_policy, so it defaulted to relaxed heuristics and kept calling the LLM 20 times with the same parameter error. Add an unconditional _looks_like_llm_provider_error check that halts auto mode with stop_reason="llm provider error" regardless of policy. Detects markers from every provider wrapper (Error calling LLM, Error calling Azure OpenAI, litellm.BadRequestError, Azure_aiException, invalid_request_error, All candidate models failed for this turn, …).

Files

mira_engine/providers/model_compat.py              (new)
mira_engine/providers/azure_openai_provider.py
mira_engine/providers/litellm_provider.py
mira_engine/providers/anthropic_provider.py
mira_engine/agent/routing.py
mira_engine/agent/research_loop.py
tests/providers/test_model_compat.py               (new)
tests/providers/test_azure_openai_provider.py
tests/test_model_routing.py
tests/test_research_loop_core.py

Test plan

• pytest tests/providers/test_model_compat.py tests/providers/test_azure_openai_provider.py — 31 passed (covers claude-opus-4-7 under every provider prefix and Azure body builder dropping temperature)
• pytest tests/test_model_routing.py — 12 passed (new: retryable-raised falls back; non-retryable-raised does not invoke fallback candidate)
• pytest tests/test_research_loop_core.py — 14 passed (new: _looks_like_llm_provider_error markers; _evaluate_continuation halts with "llm provider error" even under relaxed heuristics)
• Full regression on tests/providers/, tests/test_model_routing.py, tests/test_research_loop_core.py, tests/test_agent_loop_core.py, tests/test_agent_loop.py — 272 passed
• ruff check on newly touched files — clean (pre-existing W293/F841/I001 in unrelated lines confirmed via stash diff)
• Re-run a claude-opus-4-7 auto session against staging to confirm the surface error stays a single round and the loop halts with stop_reason="llm provider error"