Reduce reputation cost and expose empirical coverage freshness

[jehanazad]

On a live fleet, reputation evaluation repeatedly scans each node's up-to-500 trust samples for every overlapping neighbor. A 30-second profile attributed substantial processing time to these repeated scans, competing with frame ingestion.

Compute each node's trust score once per reputation pass and use that snapshot for both its own reputation and neighbor comparisons. Each new pass recomputes from current samples, so no long-lived cache can conceal new evidence. The pass also snapshots reputation membership for consistent pair enumeration.

Validation: all 519 analytics tests and all-file pre-commit checks pass. The regression checks one scan per node per pass, including a fresh scan on the next pass. This supports the real-node validation work in offworldlabs/retina-server#512.

Empirical coverage follow-up: expose last_detection_ts independently of capped per-bearing sample counts, use the recorder's range-admission multiplier when drawing evidence polygons, and return whether a calibration point was actually retained so callers cannot count rejected writes as progress. The percentile-based footprint and association gates are unchanged. Regression tests cover freshness after FIFO saturation, persistence, rejected points, and accepted evidence previously clipped by the display. All 521 analytics tests and pre-commit gates pass.