bench: add route classifier benchmark and preclassifier

[phaedrus-on-git]

Adds a labeled fast-vs-strong routing corpus, benchmark harness, and configurable keyword preclassifier for cheap high-confidence route decisions before the LLM classifier.

What changed:

• Adds test/bench/router-routing.cases.json with 20 labeled routing cases.
• Adds tools/route_classifier_bench.py for guardrail, keyword, and live OpenAI-compatible classifier comparisons.
• Adds routing.preclassifier config with target keywords, confidence threshold, and negation phrase handling.
• Keeps routing order conservative: explicit guardrail rules -> preclassifier -> LLM classifier -> default target.

Initial benchmark results:

• guardrails: 10/20, ~0ms
• keywords: 20/20, avg ~0.003ms
• live local-coder-fast classifier: 20/20, avg ~13.6s, with first cold case ~133s and warm cases mostly 6-8s

Validation:

• python3 tools/route_classifier_bench.py --cases test/bench/router-routing.cases.json --candidates guardrails,keywords
• go test ./...
• make check