scripts/category_difficulty.py refutes the IS-hardest/AL-SR-easiest priors:
DS hardest by pass@1 (47.9%, bottom-2 for 14/15 models), MI easiest (81.3%);
IS is hardest only to speed up (1.24x geomean). README finding added.

scripts/cross_pattern_transfer.py: per-category pass@1 correlates only
moderately across 15 models (mean Spearman +0.50). Clusters AL-CF +0.77,
DS-IS +0.70; MI most independent; AL best predictor of overall skill (+0.80).

modal_app/finetune_weak3.py trains QLoRA on r1-distill-qwen-7b, yi-coder-9b,
opencoder-8b (held-out excluded), merges to 16-bit, stages on the pdob-finetuned
volume. inference.py registers *-ft model keys from that volume so eval is the
unchanged pipeline.

Swap targets to the weakest fine-tune-friendly models (rescue experiment):
r1-distill-qwen-1.5b (2.8%), r1-distill-qwen-7b (26.7%), qwen2.5-coder-1.5b
(59.4%, non-reasoning control). inference.py *-ft keys synced.

Eval the 3 fine-tuned weak models on the 178 unseen held-out variants, paired
Wilcoxon vs base. Result: no positive transfer — non-reasoning qwen2.5-coder-1.5b
regresses significantly (held-out pass@1 -39/-50pp, p=0.001; hallucinated
externs, catastrophic forgetting); reasoning models nudge up off ~0 baselines
but not significantly. README finding #7.

modal_app/finetune_sweep.py: grid over epochs/lr/LoRA-rank/dropout + completion-only
loss (Unsloth train_on_responses_only) + replay data (CodeAlpaca-20k mix) to fight
the phase-1 overfitting. 2 subjects (qwen2.5-coder-1.5b regressor, r1-distill-7b)
x 7 configs; inference.py registers the *-ft variants for held-out eval.

prepare_indist_split.py holds out whole base-pattern variants (79) for a clean
in-distribution test (the old random split leaked 255/273 variants). finetune_indist.py
sweeps epochs {1,3,6,10} on the clean split to map the in-dist-transfer vs OOD-forgetting
crossover (researched recipe: lr 2e-4, alpha=2r, dropout 0.1, completion-only).

evaluate_all_modal spawns generation+CSV-write on Modal (survives --detach
disconnect, unlike evaluate_all's .map). score_modal.py scores cells on Modal
CPU; compiler.py honors PDOB_*_TIMEOUT env so broken candidates die fast.

Interrupted merges left config.json + tokenizer but no safetensors, which the
idempotency check treated as 'already merged' (so they were skipped) and vLLM
then couldn't load. Now check for safetensors and wipe+retrain partials. Add
crossover_tick.sh to idempotently drive the epoch-sweep eval->score->crossover.

The orchestrator checkpoints incrementally, so a still-generating eval CSV
looked ready and got scored on a partial (26/257 rows). Only score when all
257 in-dist+OOD rows are present; only mark DONE when the scored CSV is complete.

A prematurely-scored cell (e.g. 36 rows from a partial eval) polluted the table
with tiny-denominator garbage. Require ~257 rows or mark the cell incomplete.