Static audit of 068e\ kindling\ v2.py

• Initial-adversity streams and availability boundary. The requested fraction uses a local NumPy generator, but placement uses global torch.randperm. The condition len(available_idx) > num_to_scar silently skips a layer at equality or insufficient availability. This boundary defect is not asserted to affect the default initial call, whose scar masks begin empty and requested fraction is small. The routine returns the requested fraction, not the realized count.

• Scoring state restoration. The first scoring backward pass does not clear pre-existing gradients. The routine does not use a finally block for restoring stress and gain and does not restore its incoming train/eval mode. Fresh clones commonly avoid the first issue; reusable interfaces must not rely on that accident. Scores are accumulated absolute ambient derivatives and random regrowth is not a signed loss-descent update CheungPsy03A.

• Matching tolerance and target controls. find_iso_dose_match computes but never enforces its tolerance. The configured target list is replaced by generated targets. A pooled-percentile fallback does not establish common finite-grid support.

• Candidate and seed counts. The default iso-displacement sweep uses $11+10+9=30$ configurations. Passing all five configured adaptation learning rates gives $11+50+9=70$. The default entry point does not pass that learning-rate grid. The configured iso_dose_n_seeds=3 is unused by its one-seed iso-displacement call. The separate main routine does use its configured ten seeds.

• No longitudinal iso-displacement experiment. The iso-displacement sweeps evaluate acute outcomes and one secondary-pruning challenge. They do not run the kindling or duration-arm routines. Their comparisons cannot be presented as matched-budget longitudinal evidence.

• Dose and density omissions. The primary dose remains L1. Weight-only displacement excludes biases, gain, protection, activation, and training cost. Numerical sparsity is distinct from structural mask sparsity. Sweep displacement and acute outcomes are pre-challenge, while the stored sweep sparsity is post-challenge.

• Different secondary challenges. The main branches use a capped overall sparsity target followed by quantile pruning. The sweeps delete a fraction of surviving coordinates using kth-value thresholds. Those are not necessarily the same integer operation.

• Evaluation and efficiency semantics. The ordinary evaluator restores stress to zero, not its incoming value, and changes the model's train/eval mode. An improvement divided by dose plus $10^{-8}$ is a regularized statistic, not a physical efficiency. Smaller performance degradation does not imply higher absolute post-challenge performance.