fix(opus-4.7): flip CSV to adaptive + remove PDD_FORCE gate on github_copilot

[gltanaka]

Summary

• Flip the Anthropic claude-opus-4-7 CSV row to reasoning_type=adaptive (Anthropic enforced the new shape ~2026-05-23 17:25 UTC; PR feat(llm_invoke): add reasoning_type='adaptive' for Anthropic Opus 4.7 #1047 added the code path but deferred the CSV flip "out of scope" — this closes that).
• Teach the catalog generator about adaptive so re-generation doesn't revert the flip.
• Drop the PDD_FORCE gate on the github_copilot token-file check so non-interactive contexts (Cloud Run, library use) don't hang on litellm device-flow OAuth when no Copilot token is present.

See commit body for the full incident timeline and rationale.

Test plan

• pytest tests/test_llm_invoke.py -k github_copilot -v (three new tests pass)
• pytest tests/test_generate_model_catalog.py -k reasoning_type -v (three new generator tests pass)
• pytest tests/test_llm_invoke.py -v (no regression in existing tests)
• /gcbrun cloud-test passes
• After merge + 0.0.249 release + downstream redeploy: verify PROD fixcode returns success on POST with model=claude-opus-4-7 and no Please visit github.com/login/device warnings appear in Cloud Run logs.

🤖 Generated with Claude Code

[gltanaka]

/gcbrun

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.

[gltanaka]

LGTM — approved for merge as-is.

Verified all ten checklist items against the diff and /tmp/pdd-fix-tier2 on disk:

• CSV row: the Anthropic claude-opus-4-7 row is well-formed (all 11 columns present), reasoning_type=adaptive, max_reasoning_tokens=16000. The adaptive code path in llm_invoke.py (~L3556) does not read max_reasoning_tokens; only the budget path (~L3491) does, so the 16000 value is harmless.
• Generator self-consistency: _infer_reasoning_type returns 'adaptive' for the Anthropic row (via _is_adaptive_anthropic_model); _infer_max_reasoning_tokens returns 16000 for adaptive models. A fresh generator run reproduces the intended CSV values — no silent revert risk.
• _is_adaptive_anthropic_model correctness: matches on bare model-id claude-opus-4-7 only when provider is direct Anthropic; returns False for azure_ai/claude-opus-4-7, Vertex, and Bedrock. Future suffix variants (e.g., claude-opus-4-7-fast) would NOT match the hardcoded set — acceptable for now since the set is intentionally narrow; a follow-up can widen it when those variants exist.
• PDD_FORCE blast radius: removal is confined to the github_copilot token-file check. The interactive API-key-prompt skip at ~L2397 retains its own if os.environ.get('PDD_FORCE'): guard — unaffected.
• Backward compat: users with a valid Copilot token file still pass the check. test_github_copilot_allowed_when_token_present_no_pdd_force exercises the real branch and returns True because the token file exists, not because of any mock short-circuit.
• Warning text: pdd setup does invoke the GitHub Copilot OAuth flow via provider_manager.py. The wording is accurate.
• Test isolation: monkeypatch restores env vars after each test; tmp_path is per-test. No global model-dataframe cache in pdd.llm_invoke that would leak across tests.
• Pytest: pytest tests/test_llm_invoke.py -v → 296 passed, 0 failed; pytest tests/test_generate_model_catalog.py -v → 25 passed, 0 failed.
• Build mirror: build/lib/pdd/data/llm_model.csv does not exist — no stale mirror to update.
• Commit references: git show 8646b166e resolves; PR feat(llm_invoke): add reasoning_type='adaptive' for Anthropic Opus 4.7 #1047 grep confirms the referenced history is present.

[gltanaka]

/heal