v9.13.0

Added

• Feat: per-tool model-qualification registry (framework core, TRDD-f45eeaa0)

The single source of truth for each LLM tool's model REQUIREMENTS + its
benchmark pointer. model-qualification/registry.ts maps every LLM-using tool
to {requirements: ModelCriteria, benchmark} and exposes qualifyModelForTool().
security_scan → its real triage benchmark (973a0265); mass_scout → the
existing keyword-classification benchmark; the rest carry requirements only
(benchmark: null) until each gets a dataset. The security-triage orchestrator
now reads security_scan's requirements from the registry (real consumer).

Deliberately incremental (not premature-abstracted from N=1): per-tool
benchmark DATASETS for the other tools, the settings.yaml per-tool model map,
and generalized cross-tool selection land as each tool gets a real benchmark.

Changed

• Build: regenerate dist for the model-qualification registry