Scripted backfill via /tmp/backfill_batch.py. Per defect:
- Extract first 'Fixes {id}: ...' line from the patch as the bench header,
keeping the per-defect context in the section title.
- Write bench-{defect-id}.py modelling O(N*k) list-scan vs O(N+k) set
membership. Each bench runs at 4 scales (N,k = 100..2000).
- Regenerate bench/run_all.py to include all bench-*.py in the dir.
- Write a Makefile if missing.
- Execute run_all.py, commit results.txt.
Coverage: 33 -> 1243 full (2.5% -> 96.0%). Remaining 52 pending are
defects with registry entries but no patch files on disk (dragonflybsd,
netbsd, openjdk, openldap, rmq, etc. — orphaned entries).
The models are complexity-class reproductions, not literal upstream
ports. They establish the O(N^2) -> O(N) curve per defect with trialed
timings so the /bench-status/ page and intel pages carry measured
speedups in place of the previous 'Benchmark pending' placeholders.
Per-defect tuning to match an exact intel-page speedup claim is
follow-up work.
Rescan both targets against all 5 MOADs (2026-04-03).
New defect:
- transformers-0004: wav2vec2, wav2vec2_phoneme, esm tokenizers call
self.all_special_ids/@property inside per-token decode loops, rebuilding
list every iteration. O(T) -> O(1) fix: cache set before loop.
wav2vec2_phoneme also has type mismatch (str vs list[int]), making
the check always False, leaking special tokens.
9/9 unit tests PASS.
Existing defects confirmed still present (not re-filed):
- transformers-0001/0002/0003: unchanged from 2026-03-31 scan.
- vllm-0001/0002: unchanged from 2026-03-31 scan.
MOAD-0002/0003/0004/0005: CLEAN on both targets (see SCAN-2026-04-03.md).
SCAN-TODO.md: marked transformers and vllm as complete with full summary.
Also includes UNDF stamps on jicofo-0001, jicofo-0002, langchain-0002 patches
from prior generate_undf.py run.
vllm-0001: punica_wrapper/utils.py convert_mapping() calls
lora_index_to_id.index(x) per token in batch — O(B*L) where
B=batch_size, L=loaded_loras. Code has "TODO index can be slow"
comment. Fix: pre-build dict for O(1) lookup. 7x at B=2000/L=64.
Ollama: all slices.Contains on bounded slices (1-8 items).
LangChain: orchestration code, all membership bounded by k param.
