fix: Preserve native default reranker batch order

[buildingjoshbetter]

Scope

Default ModernBERT batch-order correction for #735, stacked on #811. Preserve model identity, weights, inference dtype, score activation, prompts, dimensions, retrieval depth and caller/server batch ceilings.

Problem and correction

The native CrossEncoder sorts the complete request before batching. Sorting separate caller-order slices instead can put a long pair in each slice. The measured mixed100 request used 64x779, 36x779 rather than the native 64x779, 36x72, adding padded work and changing native batching.

• Plan global raw-pair length order only for the known cached Alibaba-NLP/gte-reranker-modernbert-base, modern nonflattened two-string inputs, effective limit64 and more than64 pairs.
• Use the native query-plus-document length semantics. Compensate for each inner equal-length permutation and scatter scores into original occurrence positions, including duplicate pairs.
• Preserve once-per-request admission, output preflight, inference ownership, deadline checks and all batch limits. Lower explicit/adaptive limits and older/custom interfaces retain their previous bounded path.
• On a planned OOM retry, retain the failed slice, its occurrence indices and actual resolved model identity. Advance only after successful score publication.
• Keep unplanned success and retry calls compatible with positional result-writer wrappers. Existing admission regressions remain intact.

Validation

• Root focused checks and independent review passed. Independent review ran 29 tests / 137 subtests, plus 64 cases using the extracted, hash-verified ST 6.1 prediction and length bodies with stdlib fakes.
• Negative controls catch missing global order, absent/wrong inverse compensation, omitted scatter, document-only length and changed retry identity. Separate controls cover positional-only result writers on unplanned success and retry.
• Combined source fb7306ea716151f6bbef4d4e793e0ac56b45243c: 2,829 passed, 22 skipped, 3 deselected; 2,047 subtests passed, with actual sqlite-vec. Log SHA256 ddc02c30b55e3727a6eb359cc698d5a365408160048c7d699f14b95696ea922e.
• Exact isolated publication branch: 2,211 passed, 22 skipped, 3 deselected; 557 subtests passed, with actual sqlite-vec. Log SHA256 a69a832a8a99bb1303acb89bc7f3296937f12c832164088afe1dc596f1b23565.
• Configured Ruff, Python 3.10 grammar, whitespace, new-content privacy and public commit-identity checks passed. Hosted checks are pending.

An earlier combined candidate failed one existing admission test because an unplanned call passed a new keyword to its positional-only writer. The correction preserves the original interface on that path; the failed receipt and existing test are retained.

Native evidence and limits

The reviewed three-cell synthetic CUDA diagnostic completed 90 timed records, 3 warmups, 3,741 rows and 267 forwards, with verified cleanup and no OOM or timeout. It ran the combined source above; this publication branch has byte-identical model-server and microbatch test files.

All 30 default comparisons now have exact output arrays and ordered-feature parity, including all six default reranker comparisons. All 30 explicit-eight arrays exactly match their prior contiguous-eight outputs. Raw explicit-eight numerical acceptance is still 21/30, so combined release acceptance remains false.

Mixed100 reranking now matches the native 64x779, 36x72 shapes. Instrumented medians were 429.941691 ms default versus 424.593396 ms native reference: 429.941691 / 424.593396 = 1.012596275. The earlier separate run measured 631.852592 / 424.322110 = 1.489087128; this context is not a paired production p95 claim.

Reranker raw-report SHA256: 34c16142af8aa129d938f284f6eb17c42ef24e9a5a06e7fc072d5285c845affb. Diagnostic driver SHA256: 066ab7c3800fe33b7cc3fac06970cbc3b734b59652cd9e1c1d19ac953767d1e2.

The native fixture has a constant query; separate varied-query production tests cover whole-pair semantics. Serial instrumented CUDA results do not establish RPC/concurrency latency, retrieval quality, process-memory drift, MPS or thermal behavior. The old explicit-eight failures, paired runs and final soak remain open. No issue closure, main merge or release. Only synthetic test data and public model artifacts were used; no personal memories, transcripts or credentials.