chore: centralize frontier OpenAI/Anthropic chat defaults

[chrisknvidia]

Summary

• Addresses OpenClaw review feedback that the OpenAI default (gpt-5.4-mini) should be a frontier model — deferred from docs PR docs: clarify README launch overview and NVIDIA Build quickstart #33 into its own change (Fixes chore: centralize and update OpenAI/Anthropic frontier chat defaults #37).
• Centralizes chat defaults in provider_config (CHAT_DEFAULT_*, CHAT_CHEAP_OPENAI, CHAT_DEFAULT_MODELS); constants.DIMENSION_JUDGE_MODEL and llm_judge.DEFAULT_JUDGE_MODEL import from there. Harbor templates/eval.py keeps a local copy with drift tests.
• Pinned IDs (overridable via SKILL_EVAL_LLM_MODEL):
  • OpenAI: gpt-5.6-sol
  • Anthropic: claude-opus-4-8
  • Bedrock: us.anthropic.claude-opus-4-8 (AWS US geo inference profile)
  • Cheap OpenAI option: gpt-5.4-mini (CHAT_CHEAP_OPENAI)
• Normalizes bare and provider-prefixed gpt-5* IDs for both temperature compatibility and native OpenAI max_completion_tokens; raises dimension/insights judge max_tokens to 4096.

Documentation parity

• Updated the Fern provider guide, CLI reference, Tier 3 guide, and environment-variable reference with the exact provider defaults and override behavior.
• Updated the repository README so users do not have to enter the full docs site to discover a billable default-model change.
• Kept gpt-5.4-mini documented as an explicit lower-cost override rather than a silent default.

Before / after

Surface	Before	After
OpenAI default	gpt-5.4-mini	gpt-5.6-sol
Anthropic default	claude-sonnet-4-5	claude-opus-4-8
Bedrock default	us.anthropic.claude-sonnet-4-5-20250929-v1:0	us.anthropic.claude-opus-4-8
README / env reference	Exact defaults were not visible	Exact defaults and SKILL_EVAL_LLM_MODEL override are visible

Test plan

• uv run pytest tests/test_provider_config.py tests/test_llm_judge.py tests/test_oss_packaging.py tests/test_provider_agent_routing.py tests/tier3/test_judge_parse_robustness.py tests/inference/test_llm_client.py tests/validators/test_security.py -q (457 passed)
• uv run pytest -q (3,091 passed, 8 skipped, 3 deselected)
• uv run ruff check .
• uv build --python 3.13 --no-sources
• Scan source, wheel, and sdist with scripts/check_oss_boundary.py
• fern check (0 errors)
• Confirm docs contain no stale Sonnet/default-model claims

[rng1995]

The model-default centralization and temperature handling look coherent, but this branch currently has one blocking OSS-boundary CI issue.

[rng1995]

Approved after the OSS-boundary allowlist issue was fixed in 9f7bef4, the thread was resolved, and the full GitHub check matrix passed.

[rng1995]

Fresh review found two actionable issues in the OpenAI default/parameter migration. The existing OSS-boundary thread is already fixed and resolved; I will address these new findings and then update the branch from current main.

[rng1995]

Approved after fresh review and remediation. Both new P1 findings are fixed in bbbeae3, all review threads are resolved, the branch is current with main and conflict-free, local targeted/full/package/boundary verification passed, and the complete GitHub check matrix is green.

[rng1995]

Approved on final head after updating from current main. Both conflicts are resolved, all review threads remain resolved, focused verification passes (404 tests), the full suite passes (3,669 passed, 13 skipped, 3 deselected), package and source/archive OSS-boundary checks pass, all 15 GitHub checks are green, and fresh merge-tree/GitHub audits report the PR conflict-free.