Version 0.2.1 patch validation and measured costs
This patch repairs public snapshot encapsulation and response visibility. It does not change the support identity, error allowance or arithmetic proof.
Correctness defects closed
- Readonly NumPy views exposed owned model storage. Re-enabling writes could alter retained values after the exact cache was constructed. Public array getters now return independent readonly copies. The strengthened regression uses a fresh valid model per property and directly tests separated signs.
- JSONL stdout was buffered when stdin remained open. Each complete record now flushes before the next input line is read. A persistent-pipe regression requires the first answer before EOF. This is visibility, not fsync durability.
The old getter regression passed because sequential negative mass mutations
forced exact fallback and masked later value mutations. Its earlier passing
result did not establish the stronger encapsulation claim. Before-fix evidence
is retained in results/getter_defect_reproduction.json; the old runtime remains
in the historical runtime copy for reproduction only.
Fresh measurements of the patched runtime
The existing paired protocol was rerun on the patched frozen runtime: 80 container/cell/seed cases and 7,200 checked coordinate decisions, all agreeing with exact peers. No timed source changed. All historical data identities match. This reuses the previous numerical fixtures on the same machine; it is a patch qualification, not new independent data or second-machine replication.
| Input | Rows | Coordinates | Public call ms | Median speedup | Corrected bootstrap lower | 3x gate |
|---|---|---|---|---|---|---|
| ndarray | 2048 | 1 | 0.19887 | 10.64x | 8.95x | Pass |
| ndarray | 2048 | 4 | 0.45983 | 11.24x | 9.03x | Pass |
| ndarray | 8192 | 1 | 0.37337 | 23.50x | 20.49x | Pass |
| ndarray | 8192 | 4 | 1.30606 | 16.58x | 15.86x | Pass |
| list | 2048 | 1 | 0.84596 | 2.58x | 2.45x | Fail |
| list | 2048 | 4 | 1.14334 | 4.41x | 4.13x | Pass |
| list | 8192 | 1 | 3.11614 | 2.88x | 2.68x | Fail |
| list | 8192 | 4 | 4.03084 | 5.26x | 4.88x | Pass |
Comparison uses each seed's fastest measured exact comparator, including GMP direct support and integer sharp endpoints. Static preparation is amortized over 1,000 updates; already-supplied arrays/lists exclude caller conversion. Costs include input validation, fresh mass copying, predicate work and result assembly in a running process. Imports, startup, witnesses, exact margins, JSON encoding and filesystem I/O are outside the array result. Accepted filter signs retain the declared IEEE/NumPy arithmetic premise. All ordinary routes in this run were filtered; equality and adverse magnitudes remain exact fallback.
Near-boundary losses, tiny-array losses and the faster cached-endpoint path remain relevant. Output flushing is outside this in-memory benchmark. No independently engineered certified-filter comparator or production audit is claimed. The measured ratio changes from 0.2.0 are timing observations, not evidence that defensive copies or flushing accelerate the numerical hot path.
Reproduce
OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 \
python benchmarks/run_public_api.py --output benchmarks/results_public_api_v0.2.1.json
Install the declared benchmark extras first. The command must run from the full source archive; optional benchmark dependencies are not runtime requirements.
Raw report SHA-256:
6760815b0391c7d5603c4dcd24a32a2903bfadc3a6abeb2136e83c5ac9c66e1e