Stage 4C README: correct Stage 4B threshold to ep10 value (165.9)

README: add Stage 4C row + 4C/5b headline bullets; refresh Stage 4/4B to peak-checkpoint F1

infer.py: add stage_4c dispatch, pick peak-F1 epoch for student threshold

Stage 4C: ship ep10 checkpoint (peak F1 0.734 vs ep15 0.729)

Stage 4C: ship ep10 checkpoint (peak F1 0.734 vs ep15 0.729)

Stage 4B: ship ep10 checkpoint (peak F1 0.726 vs ep15 0.723)

Stage 4B: ship ep10 checkpoint (peak F1 0.726 vs ep15 0.723)

Stage 4: ship ep3 checkpoint (peak F1 0.717 vs ep10 0.710)

Stage 4: ship ep3 checkpoint (peak F1 0.717 vs ep10 0.710)

Add Stage 2b structural-pruning conversion script

Stage 4C: direct classifier-score supervision, F1 0.729 (+0.006 over 4B)

Add root infer.py dispatcher for all stages (Stage 0/1/2a/2b/4/4b)

stage_0_tight_fpr: extended-negative variant (F1 0.853 at prop-FPR 2.72 pct)

Roadmap: add stage 4b and 5b

Stage 5b: popcount reformulation, 907 gates (71 pct fewer than Stage 5), F1 0.876

Stage 4B: 15.67M student + cosine loss on 768-D, F1 0.723 (+0.013 over Stage 4)

Add Stage 2b to roadmap

Stage 2b: structural head removal (83.68M backbone, F1 0.9159 preserved)

Update roadmap table with shipped results for stages 0-5

Stage 5: Yosys synthesis (3,220 gates total for Stage 0 classifier)

Stage 4: specialist student (3.27M params, F1 0.710 vs 0.894 baseline)

Stage 3: add Stage2+Stage3 compound grid + README update

Stage 3: depth reduction results (only 1 block cleanly prunable)

Stage 2: attention-head pruning results + mask + apply_mask.py

stage_0: add prop-specificity audit + note in README

Remove stray __pycache__ artifact

Stage 1: add classifier.safetensors

Stage 1: output-channel pruning (Path A, exact parity with Stage 0)

Add Stage 0 classifier weights

Add YAML metadata frontmatter

Initialize repo with Stage 0 baseline

initial commit