fix(profiles): gate DeepSeek thinking-off default on reasoning-hint compatibility

[himorishige]

Fixes the tier-call half of #121.

What

apply_deepseek_overrides() unconditionally injects the vLLM-only chat_template_kwargs.enable_thinking=false hint for any target whose model id contains deepseek-v4. Providers with strict request validation (observed: Fireworks AI serverless) reject the unknown field with HTTP 400 Extra inputs are not permitted, breaking every routed call for such tiers.

This PR gates the body default on model_accepts_reasoning_hint() — the same model-id compatibility check LLMClassifierConfig.disable_reasoning auto-detect already applies before sending the identical hint on classifier calls. The provider-neutral X-Inference-Priority header default is unchanged, and explicitly provided extra_body still wins as before.

Pairs with #123 (adds fireworks to the deny list); together they stop the 400 for Fireworks-served deepseek-v4 targets.

Validation

• New unit test: deny-listed model id keeps extra_body untouched while still receiving the batch-priority header; existing TestDeepSeekOverrides cases unchanged and passing.
• tests/test_deterministic_routing_profile.py + tests/test_reasoning_hint.py: 47 passed. Two failures (TestProfileStructure::test_shared_stats_accumulator, test_overflow_reroutes_to_custom_strong_id_through_full_profile) are pre-existing on unmodified main in my environment (same class of stale-build failure noted in feat(serve): route-log per-request routing stats for profile-config serve #118's validation notes).

Summary by CodeRabbit

• Bug Fixes

  • Improved compatibility for DeepSeek models that do not support reasoning hints.
  • Prevented unsupported “thinking off” request parameters that could cause HTTP 400 errors.
  • Preserved batch inference prioritization for affected requests.

• Tests

  • Added coverage to verify incompatible DeepSeek models receive valid request settings.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: b363c6d1-1f92-4b9a-a786-38228865ac43

📥 Commits

Reviewing files that changed from the base of the PR and between 060ad75 and 15e8ec1.

📒 Files selected for processing (2)

• switchyard/lib/profiles/tier_target_builders.py
• tests/test_deterministic_routing_profile.py

---

Walkthrough

DeepSeek v4 thinking-disabled overrides are now gated by model compatibility with reasoning hints, and a regression test covers incompatible model IDs while preserving batch-priority routing.

Changes

DeepSeek override compatibility

Layer / File(s)	Summary
Reasoning-hint gating and regression coverage switchyard/lib/profiles/tier_target_builders.py, tests/test_deterministic_routing_profile.py	apply_deepseek_overrides checks model_accepts_reasoning_hint before adding the DeepSeek v4 thinking override; the test verifies incompatible models omit extra_body while retaining the batch-priority header.

Estimated code review effort: 2 (Simple) | ~10 minutes

Poem

I’m a rabbit with code in my paws,
Thinking stays off when hints break the laws.
Batch headers hop neatly in line,
No extra body slips through the design.
Sniff, test, and celebrate—fine!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: gating DeepSeek's thinking-off default on reasoning-hint compatibility.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[elyasmnvidian]

Nice — gating the tier default on model_accepts_reasoning_hint matches what the classifier already does, so both paths now agree on when to send enable_thinking=false. (For context: apply_deepseek_overrides sets default request options for DeepSeek tier calls, and this hint is one of them.)

One thing to flag for later: both the tier default (this function) and the classifier default now depend on that one check being complete. The check works by looking for provider names inside the model id, so a new backend that rejects the hint will still receive it until someone remembers to add its name to _NO_REASONING_HINT_TAGS. That is easy to forget, and the failure shows up as a hard 400 in production.

Concrete suggestion: consider deciding this on the target/endpoint config rather than the model name. For example, add an optional accepts_reasoning_hint flag to the target (LlmTarget in llm_target.py), default it from model_accepts_reasoning_hint(model) when the config doesn't set it, and have both apply_deepseek_overrides and the classifier presets read that flag. Then a strict backend can be marked incompatible in its own config, with no code change and no name list to keep in sync.