v0.8.0

0.8.0 (2026-09-07)

Features

• analysis: AttentionRegime and the attention decision rule (#432) (6e9561a), closes #403
• analysis: saliency shape statistics for the attention-regime diagnosis (#428) (8fc0f25), closes #402
• benchmarks: bnnr_sh and bnnr_diagnosis conditions for the search policies (#446) (52a602b), closes #415
• config: diagnosis thresholds as config with no defaults (#433) (d08911b), closes #405
• reporting: surface the attention evidence and promote ECE into the table (#439) (9f33dc8), closes #412
• training: CandidateSelector protocol and the SELECTORS registry (#431) (75c949e)
• training: fall back to metric_argmax below min_confidence (#445) (6972733), closes #414
• training: hard_quantile_acc and robustness_gap (#430) (ce98980), closes #404
• training: indistinguishability test for candidate selection (#437) (b59b4a9), closes #408
• training: mandatory run-record fields for selection and compute (#434) (565a0d1), closes #410
• training: search_policy with exhaustive, diagnosis_single and successive_halving (#440) (97c80db)
• training: shadow mode, record the diagnosis evidence without acting on it (#435) (1b4656c)

Bug Fixes

• augmentations: make ChurchNoise the same transform on both paths (#423) (7241229), closes #396
• augmentations: make DifPresets and ProCAM the same transform on both paths (#429) (34b9799), closes #421
• benchmarks: extract stats.py, fix rank-biserial sign and paired bootstrap FIX-0-5 (#441) (61b2536)
• config: deprecate the guessed XAI knobs and stop double-counting accuracy (#436) (6d9b0e3), closes #409
• selection: scale candidate differences by measurement noise, not by spread (#438) (7ced1f1), closes #407

Documentation

• describe what ICD and AICD actually mask, and when each applies (#424) (ff8af15), closes #397