fix(providers): self-heal temperature rejection on the OpenAI-compatible path

[lorenzozanee]

Summary

• Extend the Anthropic temperature self-heal (classify the 400, retry once, remember the model) to the generic OpenAI-compatible branch (ChatOpenAIWithReasoning), so requests to Anthropic models through OPENAI_BASE_URL recover from "temperature is deprecated for this model." automatically.
• No change for models that accept temperature — the configured value keeps being sent.

Why

Next-generation Claude models (claude-opus-4-8, claude-sonnet-5, ...) reject any request carrying temperature with HTTP 400. At v0.1.14 the native Anthropic Messages path detects the error, retries without the field and remembers the model, but the generic OpenAI-compatible branch still sends temperature unconditionally, so running LANGCHAIN_PROVIDER=openai with OPENAI_BASE_URL=https://api.anthropic.com/v1/ fails on every call.

Closes #1223

Changes

• ChatOpenAIWithReasoning._remember_temperature_unsupported classifies the 400 with the existing _is_anthropic_temperature_unsupported_error detector, logs once, and records the model in _ANTHROPIC_TEMPERATURE_UNSUPPORTED — the same set the native path uses.
• _generate / _agenerate retry once without the field; _stream / _astream do the same (the 400 is raised before the first chunk, so the retry cannot duplicate output). The async stream_usage fallback moves to a named _astream_with_usage_fallback to mirror the sync _stream layering — no behavior change.
• _get_request_payload omits temperature up front for remembered models, so later calls skip the doomed request.

Test Plan

• Existing tests pass (pytest --ignore=agent/tests/e2e_backtest --tb=short -q — verified for the touched components: agent/tests/test_llm.py, agent/tests/test_provider_header_isolation.py, agent/tests/test_env_schema.py)
• New tests added: mocked-error regression tests in agent/tests/test_llm.py covering invoke self-heal and the remembered-model fast path, supported-model preservation, stream self-heal, and async invoke; all fail on the pre-fix code with the verbatim 400
• Tested manually (provider-dependent behavior verified with mocked 400/SSE responses via httpx.MockTransport, no live credentials)

Checklist

• No changes to protected areas (src/agent/, src/session/, src/providers/) without prior discussion — direction raised and left open in temperature sent unconditionally on the OpenAI-compatible path to Anthropic (native-path fix at v0.1.14 doesn't cover it) #1223; follows the self-heal precedent from Swarm run.json under-reports output tokens ~18-36x: streamed calls never request usage, and the fallback ignores tool-call arguments #1224 on the same file
• No hardcoded values (API keys, file paths, magic numbers)
• Code follows CONTRIBUTING.md guidelines
• Documentation updated (if user-facing change) — no doc change needed; behavior now matches the native path

[warren618]

Merged, and I was wrong to decline this direction on #1223.

My argument there was that the generic OpenAI-compatible branch is "deliberately provider-agnostic" and teaching it Anthropic-specific error prose would couple it to every provider quirk it currently stays clean of. That is not true of the branch as it stands: _STREAM_USAGE_UNSUPPORTED in the same file is already a classify-the-error, retry-once, remember-the-model self-heal for a provider quirk, inside ChatOpenAIWithReasoning, landed for #1224. Same class, same file, same branch. You said as much in your checklist and you were right. Sorry for the round trip.

One hardening pushed on top (3e27d25a). The stream retry restarted the whole stream on the 400, with a comment that the error arrives before the first chunk so output cannot duplicate. That is true, but it is a property of the provider, not of this code — the retry now re-raises once a chunk has been emitted, so a mid-stream error matching the detector can never replay already-emitted output.

The same commit lands the two things I promised on #1223 and had not done: ANTHROPIC_API_KEY joins env_schema (it was read through a bare os.getenv, so provider doctor and preflight could not tell a missing key from a working one), and .env.example now states that pointing OPENAI_BASE_URL at Anthropic's compatibility endpoint is not a supported route.