Root Cause

brick_scores_from_profiler() in gguf.rs hardcodes all scoring fields:

• actual_us = per_token_us (total_ns / profiler.total_tokens where total_tokens=952K brick elements, not ~3K decoded tokens)
• budget_us = same as actual (self-referential)
• score: 100, grade: "R" (not even a valid grade), gap_factor: 1.0

Impact

JSON output reports all bricks as perfect. LmHead shows 1.9µs instead of real 595µs.

Fix

• Use stats.avg_us() for actual_us
• Derive decoded_tokens from LmHead.count
• Compute real score/grade/gap via compute_brick_score()

Refs GH-420 (parent)