feat: in-session model + effort routing advisor
measure.py route --task "..." recommends both a model and a reasoning effort for a task, in the current platform's own terms. Very cheap models (haiku, sol, luna) and low/minimal effort are reachable only for easy tasks — a hard floor no significant task can bypass.
- Two-axis: model AND effort, together.
- Native per platform: Claude, Codex, OpenCode, Copilot, Hermes (only the current one is used).
- The installed model-routing guidance block is now live, two-axis, and platform-native.
- Local, deterministic, fail-open; classification input capped; never raises.
QA: floor proven across ~4,076 fuzz cases + structural argument; security review clean (no ReDoS/eval/injection); 324 tests pass.