[kleedaisuki]

Purpose

Per-worker routing now fits contention using observed overlap rather than selection-time peer counts. The persisted sufficient statistics retain the same four algebraic features but a changed label meaning, so old v0.6.0 priors must not be loaded under the new semantics.

Changes

• Use a v2 learning-key namespace while keeping the model-store schema/feature version unchanged, preserving old rows and writeability.
• Display the effective model identity in human and machine-readable summaries; document one intentional cold-learning run after upgrade.
• Add a cross-run CLI regression: seed legacy key, remove v2 key, verify zero old-prior hits, legacy state retention, v2 persistence and later reuse. Keep fixtures inside .temp.

Verification

• Windows MSVC Debug full CTest: 27/27 passed.
• Physical Release CUDA binary on existing v0.6.0 model.db: first run had 0 v2 prior hits, second had 3; database contained 4 old v1 rows and 1 new v2 row.
• Cross-platform GitHub Actions pending.

Limitation

The first run after upgrade starts new model priors. This is deliberate; old data is retained but not mixed with differently defined labels.