feat: route on a calibrated classifier, live load, and measured quality

[Yash-Chindam]

Fourth PR closing gaps between the design spec and the implementation.

Gap this closes

§7.2: "Use a lightweight classifier or calibrated model for task and complexity prediction", plus the routing features it lists that were missing or static: structured-output requirement, current queue delay, GPU capacity, and historical quality by task and model.

What changed

• classifier.py: multinomial naive Bayes over unigrams and bigrams, trained at start-up from config/routing/task-classifier-v1.jsonl. No new dependency. Posteriors are temperature-scaled on a held-out split; a prediction under 0.5 abstains to general. A declared routing.task is never overridden. Keyword rules remain the fallback when the dataset is absent.
• Complexity (low/medium/high) is predicted by a second head and pulls a request toward higher measured quality.
• routing.structured is now a hard filter against supports_structured_output on the model card.
• Quality is looked up per task from benchmark history (BenchmarkRun.task), falling back to the card's headline figure.
• load.py: observed queue delay (EWMA per model) replaces the catalog estimate; engine saturation from the last metrics scrape is a penalty on local models, discarded after 30 seconds.
• The response and the trace report task_source, task_confidence, and complexity.

Limits worth knowing

• The training and evaluation prompts were written by hand, in one voice, for this PR. The held-out result (42/42 tasks, 88% complexity, ECE 0.035) checks that the mechanism works; it is not evidence of accuracy on real traffic.
• One gateway faces one engine, so saturation costs all local models equally. It only affects local versus external.
• Required modality is not implemented; message content is text only.
• The only benchmark run in the catalog covers extraction, so quality history changes nothing in the committed configuration yet.

Test plan

• 35 new tests (classifier accuracy and calibration on unseen prompts, abstention, structured hard filter, per-task quality, queue feedback, saturation and its privacy limit)
• ruff format --check ., ruff check ., mypy clean
• pytest tests/unit tests/integration: 241 passed, coverage 98%
• Playwright end-to-end suite: 9 passed locally against the running gateway

🤖 Generated with Claude Code