Selected upstream review, 12 October 2026¶

This review covers 210 commits, including 128 non-merge commits, from upstream 68804629 (0.4.2) through c73908d (0.4.6). The comparison records the boundary. Laya-Pro remains version 0.3.29 with selected upstream changes; it keeps its English fallback, typed schema identity and decision retention policies. Checkpoint weights and shipped calibration files are unchanged.

Changes adopted¶

Additional fixes¶

Python and TypeScript embedding caches now track a clear generation. An embedding started before a clear can return its own snapshot, but cannot repopulate a cache belonging to a replacement model. Locking the dimension check alone did not fix that race.

HTTP tournament eligibility also controls the raised option cap. An ordered choice that requires the one-pass path retains its 100-option cap and is refused before inference when oversized. Generated elimination calls contain at most 64 questions. MCP route deadlines and batch sizes reject booleans instead of interpreting False as an unset control.

Independent review found that generated tournament question ids could change automatic workflow routing between rounds. Python and TypeScript now resolve the original schema first and keep its checkpoint and detected language throughout elimination. The review also caught two test errors in the API annotation check and cached batch result shape; both are corrected.

Measurements¶

The calibration comparison uses seeded synthetic rows, CPU Torch with two threads, one warmup and three alternating before/after repeats. Mixed 2/9-option argmax rows fitted T=1.0 before and T=0.5 after; true unpadded NLL fell from 0.765155 to 0.479609. Mixed negative 2/4-option rows fitted T=5.0 before and T=0.5 after, with NLL 0.881463 to 0.241116. A fixed-width interior optimum stayed at T=0.522357. These demonstrate fitter correctness; they do not establish pretrained accuracy or justify replacing deployed calibration files.

Fitter medians for the mixed 2/9 and negative 2/4 panels were 16.179→2.285 ms and 61.440→1.307 ms. The interior mixed 3/20 panel cost 42.816→46.518 ms and changed NLL only from 1.063982619 to 1.063981773; the fixed-width panel cost 50.778→51.049 ms. The fix adds width-group work when both fits already converge. These are small fitter-only measurements, separate from inference latency. Harness, raw comparison.

The native TypeScript embedding benchmark uses the retained English checkpoint, a 20-label catalog and two distinct requests per repeat, two warmups and three alternating timed repeats. Reusing label vectors reduced measured encoder rows from 126 to 6, with identical rankings, and median request cost from 294.525 to 33.726 ms. This measures embedding and ranking, excluding decision inference. Two ONNX Runtime threads, a 64-token embedding window and graph hashes are recorded with the raw samples.

The HTTP English smoke used checkpoint 55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851, two CPU Torch threads, two warmups and three alternating repeats. One-pass cost was 344.143 ms, versus 584.469 ms for HTTP tournament inference on the same 20-label request with groups of 16. HTTP timing includes the transport; this demonstrates additional cost, not a speed claim. Finalist probabilities matched direct Python tournament output exactly, and a 128-label request completed with eight finalists. The checkpoint's 11+ temperature is clamped to 0.5 by the existing loader and remains uncalibrated. Measurements ran on an active Windows workstation. Harness, raw samples.

Native TypeScript passed 887 tests, type checks and an external packed-package consumer. The HTTP client passed 33 tests, including packaging and typed tournament metadata. Four real English tournament scenarios covered zero, one and two elimination rounds. Winners, finalists, call counts and usage matched Python exactly; the largest returned numeric difference was 1e-4, one four-decimal rounding unit. Parity record.

Validation¶

The shared 92-suite Python gate was exercised. Its pytest group passed 546 tests with one checkpoint-dependent skip; after correcting a new mock's label set, the shortlist script passed 249 checks on rerun. The API contract passed 688 checks. Separate evaluation suites passed 237 tests, with 50 API and 40 ONNX runner checks. Ruff and compilation passed. The strict clean docs build reported no issues, and identity/crawl checks covered 49 sitemap URLs. Follow-up independent review verified all three corrections and found no further actionable defects in the complete diff, including new files.

Deferred changes and limits¶

The new Go port currently supplies language detection and internal tokenizer foundations; its agent, router and ONNX inference layers are planned for later changes. It is not adopted as a production decision SDK. The upstream multilingual default would change ambiguous-state routing, so Laya-Pro retains English and supports explicit multilingual configuration.

Native TypeScript hook deadlines remain deferred. The newer formatting and timer cleanup do not prevent an abandoned callback from modifying its shared context. Awaited hooks keep their existing contract. Docker and Nix are unavailable on this Windows host. Docker build and runtime checks run in CI; Nix validation remains unverified. ONNX evaluation revisions identify configuration/tokenizer source, not a hash of separately supplied graphs.

Tournament probabilities and gates describe the surviving finalists. Elimination can change the selected label, costs additional inference and needs workload-specific quality evaluation. No new accuracy, calibration or cross-language quality claim is made.