CRITICAL — acdc.py: pruned_so_far passed all tested edges (retained+pruned)
instead of only pruned ones, breaking ACDC's greedy edge-removal algorithm.
Correct: patch only previously pruned edges + current candidate.

HIGH — multi_arch.py: adjust_attributions_for_gqa() crashed with ValueError
when unpacking head_scores (Dict[int,float]) as (layer_key,head_key) tuples.
Fixed to correct per-layer key assembly: adjusted[(layer, head_key)].

CRITICAL — cross_model.py: bin index int(norm/0.1) returns 10 for norm=1.0,
out-of-bounds for 10×10 grid. Clamped to min(int(norm/bin_size), max_bin)
in _jaccard, _attribution_pearsonr, and _find_consensus_heads.

CRITICAL — cross_model.py: empty union in _jaccard returned 1.0 (identical).
No evidence of overlap → correct default is 0.0.

HIGH — cross_model.py: unique_to_a/b subtracted bin-count from position-count,
mixing granularities. Fixed to bin-set difference: len(bins_a - bins_b).

CRITICAL — steering.py: _VERSION hardcoded as "3.6.0". All steering vector
HTML reports showed wrong package version. Applied importlib.metadata lookup
matching evidence_vault.py fix from v4.2.1.

HIGH — README.md: version badge and PyPI example updated 4.2.0 → 4.2.3.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>