feat(models): route normal to DeepSeek V4 Flash, classifier to gpt-oss-120b

[winlp4ever]

What

Alternative to the GLM base in #268, which regressed in testing: GLM 5.3 Flash via OpenRouter did not emit real tool_calls (the agent hallucinated the whole tool loop as text, so nothing hit the board) and was too slow for the 2s classifier timeout (it took ~2030ms, so the classifier always fell back to medium and complex routing died).

DeepSeek V4 Flash tests clean for tool calling. This routes:

• normal (lite) -> DeepSeek V4 Flash (deepseek/deepseek-v4-flash:nitro, OpenRouter) when an OpenRouter key is present; falls through to gpt-5.4-mini with no key.
• complex (pro) -> gpt-5.5 (unchanged).
• classifier -> gpt-oss-120b (openai/gpt-oss-120b): open-weight, native structured output, ~$0.03/$0.17 per M tokens. Pinned by id so it stays independent of the base model.

How

• models.yml: deepseek-v4-flash becomes the first lite entry (auto base); add gpt-oss-120b (lite, no :nitro since the openai/ route is single-provider). gpt-5.5 stays the first entry overall, so the no-tier fallback and complex default are unchanged (the no-tier-default lesson from feat(models): route normal to GLM 5.3 Flash, complex to gpt-5.4 #268's review).
• catalog.resolved_by_id(id): resolve a specific model by canonical id.
• auto_model.py: pin the classifier to gpt-oss-120b, falling back to the lite default then any model. Bump AUTO_MODEL_TIMEOUT_SECONDS 2 -> 5 for OpenRouter first-token latency headroom.

BYOK fallback

DeepSeek and gpt-oss are OpenRouter-only, so with no OpenRouter key the base falls to gpt-5.4-mini and the classifier to the lite default, both native OpenAI. No one is stranded.

Test plan

• ruff check topix test/unit: clean
• pytest test/unit: 706 passed (added resolved_by_id tests)

Notes

• feat(models): route normal to GLM 5.3 Flash, complex to gpt-5.4 #268 (GLM base) stays open for separate testing; this is the alternative to merge if DeepSeek wins.
• The signed-out BYOK default is out of scope here (handled in feat(models): route normal to GLM 5.3 Flash, complex to gpt-5.4 #268).