Summary

apr cbtop --model-path <FILE> --headless --json produces completely fake BrickScore data in JSON output, despite the BrickProfiler collecting correct GPU timing internally. The stderr shows real data (LmHead avg=595µs) but JSON reports LmHead actual_us=1.9µs.

Root Cause

brick_scores_from_profiler() in gguf.rs hardcodes all scoring fields and uses wrong denominator.

Bugs (5 total)

Bug 1: gguf.rs:406-413 — brick_scores_from_profiler hardcodes everything

• score: 100, grade: "R", gap_factor: 1.0 all hardcoded
• budget_us and actual_us both set to per_token_us (total_ns / profiler.total_tokens)
• profiler.total_tokens() counts brick ELEMENTS (~952K), not decoded tokens (~3K)
• Result: all bricks show score=100, gap=1.0, actual=budget — completely useless

Bug 2: cbtop_measure_batch.rs:328-335 — weighted score truncation

• pmat_brick_score uses 7 hardcoded weights [1.5, 6.0, 1.0, 10.0, 3.5, 1.5, 12.2]
• Real profiler returns 11 bricks — zip silently drops 4
• Aggregate brick_score wrong

Bug 3: cbtop_measure_batch.rs:362-365 — hardcoded PMAT scores

• rust_project_score: 173.9, tdg_score: 98.1, cuda_tdg_score: 95.2 are constants
• Not computed from any real data

Bug 4: cbtop_measure_batch.rs:369-374 — hardcoded FalsificationSummary

• total_points: 137, passed: 137, failed: 0, blocked: 0
• No actual falsification tests run

Bug 5: cbtop_measure_batch.rs:338 — hardcoded target

• target_tok_s = 976.0 is a stale CPU spec target
• Should be derived from hardware capability or removed

Contract Violation

Violates C-GDP-001 (gpu-decode-profiling-v1.yaml): brick data must reflect real GPU execution time.

Fix Plan

1. Use stats.avg_us() for actual_us (per-call GPU time)
2. Derive decoded_tokens from LmHead.count (exactly 1 per decoded token)
3. Compute real score/grade/gap_factor using existing compute_brick_score()
4. Use equal weights for dynamic brick count instead of hardcoded 7-element array
5. Remove hardcoded PMAT scores and falsification summary — report 0 if not computed