recipe(ms-marco-MiniLM-L12-v2): add CPU reranking recipes

[ssss141414]

Summary

This adds CPU fp32 and fp16 reranking recipes for cross-encoder/ms-marco-MiniLM-L12-v2, a cross-encoder that scores query-document pairs for ranking. The contribution has Effort L2 and Outcome L2 because its independent delta remains recipe-only while generic L2 reranking support is owned by stacked dependency PR #1322. The highest Goal verdict reached is L3 PASS with full coverage across both required CPU precision tuples.

Model metadata

What the model does

Cross-encoder reranker that scores a query-document pair with a single scalar relevance logit for ranking. Evidence: Hugging Face model metadata pipeline tag and model card task tags; pinned checkpoint configuration identifies BertForSequenceClassification with num_labels=1. Confidence: verified.

Primary user stories

• A user supplies a search query and a small candidate document list to obtain relevance scores used to rank results. Evidence: MS MARCO reranking model card intent. Confidence: mapped.
• A retrieval pipeline supplies top-k candidates to obtain better ordering before final answer generation. Evidence: cross-encoder reranker usage pattern. Confidence: inferred.

Supported tasks

• text-ranking across the checkpoint, Transformers, Optimum ONNX, and WinML support surfaces. Evidence: pipeline tag, WinML inspect pipeline task, and Optimum probe tasks. Confidence: mapped.

Model architecture

BertForSequenceClassification
|-- BertEmbeddings
|-- BertEncoder (12 Transformer layers)
|   |-- Self-attention
|   |-- Feed-forward
|   `-- Residual + LayerNorm
`-- Sequence classification head (single scalar logit)

• Source/confidence: pinned checkpoint cross-encoder/ms-marco-MiniLM-L12-v2 configuration and Transformers BERT class structure (mapped).

Validation and support evidence

Baseline

Current main and the frozen baseline remain 0876e5ae1c98a169a6137e092e0d7b30bf9cee33 with WinML 0.0.1.dev0. Recipe-free auto-resolution used text-classification: build PASS in 49.9s, followed by perf PASS at 98.69 ms mean, 98.21 ms p50, 10.13 samples/s, and +79.4 MB RAM. The build output reported an Optimize stage despite --no-optimize. Canonical text-ranking config was unsupported by TasksManager (exit 2), and canonical text-ranking eval was unsupported (exit 1). The Optimum probe was VENDOR-ONLY: BERT export was vendor-covered for text-classification, with no task added by WinML at baseline.

The dependency baseline for this stacked contribution is PR #1322 at 3708969b731425b0c6d4b97920d1b5e6519bb013, advanced by fast-forward from 030e79fa6482acb08e51338ab4172ac4387b387d. The current L12 candidate is 8dae90fbd75c30c79d0bd5ec60f7817d23d5bd6d, with that exact dependency commit as its parent.

Goal

• Effort: L2.
• Goal ceiling: L3.
• Outcome: L2.
• Success definition: L0 build/structure, L1 performance, L2 numerical parity, and L3 bounded reranking functional smoke.
• Ceiling change: none; the committed L3 ceiling was reached.

The L2 context is stacked: PR #1322 owns generic reranking task resolution, inference/evaluator dispatch, grouped ranking data handling, and ranking metrics. This independent commit adds model recipes only and does not claim to add generic source support.

Outcome

L3 PASS was reached with full coverage. Both required tuples, CPUExecutionProvider/cpu/fp32 and CPUExecutionProvider/cpu/fp16, pass L0, L1, and L2; the bounded fp32 public evaluator smoke passes L3. Deferred tuples: []. Unresolved tuples: []. The shipped Outcome remains L2 because this PR's independent delta contains exactly two CPU recipes and relies on PR #1322 for generic L2 reranking code.

Learner findings v4 retain bert-011, bert-013, bert-014, bert-015, and bert-016, and update:

• bert-012: L12 reranker CPU fp16 completes a bounded one-run probe despite historical multi-iteration timeouts. The current tuple is PASS; scaled-run completion and a stable latency distribution remain unmeasured.

The methodology audit is not a no-friction result: reviewer R4 exposed a hash-valid root with contradictory terminal-state semantics, so learner finding _meta-111 now requires semantic agreement across terminal-state and identity fields in addition to immutable roots and hashes. Lane A pushed the implementation in 2c5700a047fc85e5fb3405d361f8fb06e1c846fd and exact-SHA metadata finalization in e70be83aefd621dbb4170aa647fe6344a6d4884c; review it at https://github.com/gim-home/ModelKitArtifacts/compare/main...ssss141414/learner-l12-bert-knowledge-20260823?expand=1. These knowledge and skill edits remain separate from this model PR.

Per-EP/device/precision results and Functional smoke Eval

Tier	EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM / RSS	Detail
L0	CPUExecutionProvider / cpu	fp32	PASS	-	-	-	-	133,589,610 bytes; 204 FLOAT + 56 INT64 initializers; 388 nodes; three [1,512] int32 inputs; [1,1] float logits
L0	CPUExecutionProvider / cpu	fp16	PASS	-	-	-	-	66,856,699 bytes; 204 FLOAT16 + 56 INT64 initializers; 389 nodes; three [1,512] int32 inputs; [1,1] float logits
L1	CPUExecutionProvider / cpu	fp32	PASS	101.383 ms	99.193 ms	9.86 samples/s	+79.65 MB	30 iterations after 5 warmups; reused after recipe byte-equivalence and Eval-only dependency delta
L1	CPUExecutionProvider / cpu	fp16	PASS	187.938499962911 ms	-	5.320889547364412 samples/s	647761920 bytes before; 695275520 bytes after; +47513600 bytes	True-fp16 model; 1 measured iteration, 0 warmups, 1800-second cap; no peak RSS was sampled and no peak claim is made

Historical context only: prior candidate 07a7ee4fa097210a43e2ca9ce7f68b659a2cedb0 exceeded 2700 seconds at 2743.499 seconds for a 30-iteration/5-warmup run, and a 10-iteration/3-warmup attempt reached 600 seconds after setup without a result. Those bounded prior attempts are superseded as the tuple verdict by the current one-sample PASS and are not stable benchmark evidence.

R4 terminal-state integrity is repaired in a fresh immutable Tester v3 root: the corrected controller records result PASS, phase completed, and outer exit 0. Cross-file validation reports semantic agreement, and the 26-entry final seal reports zero hash mismatches. The old sealed root is preserved only as incident provenance and is not authoritative for the current fp16 result.

L2 named-input parity PASS for both artifacts over three pairs: fp32 cosine 0.9999999999999853, max absolute difference 0.000002384185791015625; fp16 cosine 0.9999990532003638, max absolute difference 0.03248405456542969. L0 and L2 were reused only after range-diff and blob comparison proved both recipes byte-identical across the restack; the dependency delta is Eval-only.

Functional smoke Eval

L3 PASS on final candidate 8dae90fbd75c30c79d0bd5ec60f7817d23d5bd6d, FP32 CPUExecutionProvider/cpu, through the public winml eval command and winml.modelkit.eval.reranking_evaluator.WinMLRerankingEvaluator. The pinned dataset is mteb/scidocs-reranking revision 56a6d0140cf6356659e2a7c1413286a774468d44, split test, streamed without shuffle. The deterministic first 2 groups were selected and processed with 0 skipped groups, producing 20 query-document pairs. Positive and negative passage text was materialized before scoring; positives are relevant, negatives are non-relevant, and each pair produces one scalar relevance logit ranked in descending order. Caps were 10 candidates per group and sequence length 512.

Results: MRR@10 1.0, Recall@1 1.0, and Recall@10 1.0; 2 scored groups and 0 groups without a positive. This is a bounded functional operability smoke only, not representative accuracy or benchmark quality, and it does not claim fp16 Eval accuracy. Baseline main could not run canonical text-ranking Eval; PR #1322 at 3708969b731425b0c6d4b97920d1b5e6519bb013 supplies the generic public evaluator capability exercised here.

Delta

The independent PR #1334 delta adds exactly:

• examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/reranking_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/reranking_fp16_config.json

There are no modified or deleted files and no source, test, or README changes; the production recipe README remains untouched. The canonical baseline recipe is NOT-COMPARABLE: baseline TasksManager could not emit canonical text-ranking config, so there are no baseline JSON-pointer old/new values. Both recipe blobs are byte-identical to old candidate 07a7ee4fa097210a43e2ca9ce7f68b659a2cedb0 after restacking on dependency 3708969b731425b0c6d4b97920d1b5e6519bb013. Recipe-free acceptance PASSes with resolved task reranking, AutoModelForSequenceClassification, and canonical paired BERT inputs. The delta remains reducibility-consistent with the charter.

Bug fix explanation (dependency PR #1322)

1. Symptom and trigger: public no-dataset winml eval --task reranking --samples 2 selected an ID-only MS MARCO dataset whose rows lacked candidate passage text, so WinMLRerankingEvaluator rejected them before scoring.
2. Root cause: the default dataset contract and evaluator schema were incompatible. Resolving IDs through the canonical 8.8M-passage corpus or top1000 archive would require an unbounded multi-gigabyte download or scan even for a tiny smoke.
3. Changed symbols and mechanism: PR Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322 changes _DEFAULT_DATASETS['reranking'], _RERANKING_SCHEMA, RerankingDatasetMode, detect_reranking_dataset_mode, and WinMLRerankingEvaluator group materialization. The default is pinned streaming SciDocs; grouped query/positive/negative text is detected and materialized with deterministic candidate IDs, positive retention, and max_candidates=10.
4. General rule: the fix is schema- and data-driven for grouped reranking datasets, with no checkpoint-specific branch or hardcoded L12 behavior.
5. Compatibility and blast radius: existing pairwise and already-materialized candidates_column paths remain supported; materialized candidate order remains unchanged and is not capped. The intentional changes are the default streaming SciDocs dataset and bounded grouped positive/negative text support.
6. Regression evidence: focused default/evaluator coverage passed 109 tests; commands/config/build/compiler/session/eval passed 3641 tests with 9 skipped and 1 warning; datasets passed 78; utils passed 279 with 1 skipped; Ruff passed; mypy reported no issues in 438 source files. The final candidate public CLI smoke passed on 2 groups and 20 pairs with the metrics above.

This generic bug fix belongs to dependency PR #1322. PR #1334's own diff remains exactly the two recipes listed above.

Analyze summary — component level and op level

ANALYZE-PARTIAL-SUCCESS (exit 1): OpenVINO CPU rule analysis completed, while CPUExecutionProvider has no runtime rule data. This is static compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	embeddings; 12x encoder attention/FFN; scalar head	429 mapped, 7 unmapped; mapped with explicit optimization provenance	None

Final optimization removes hierarchy metadata and renames nodes; semantic components use the preserved tagged export hierarchy or native ONNX scopes, while op/EP counts come from the final graph.

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	388 operators / 16 unique types	Reshape 120; Gemm 74; Transpose 48; Add 38	OpenVINOExecutionProvider/CPU fully supported

CPUExecutionProvider/CPU has no rule data and therefore an empty classification; this is not a runtime support claim.

Reproduce commands

$OUT='temp/ms-marco-MiniLM-L12-v2-repro'
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/reranking_fp32_config.json -m cross-encoder/ms-marco-MiniLM-L12-v2 -o $OUT/fp32
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/reranking_fp16_config.json -m cross-encoder/ms-marco-MiniLM-L12-v2 -o $OUT/fp16 --precision fp16
winml analyze --model $OUT/fp32/model.onnx --ep all --output $OUT/analyze.json
winml perf -m $OUT/fp32/model.onnx --device cpu --ep cpu --iterations 30 --warmup 5
winml eval -m $OUT/fp32/model.onnx --model-id cross-encoder/ms-marco-MiniLM-L12-v2 --task reranking --ep cpu --device cpu --samples 2

[ssss141414]

REQUEST_CHANGES

Reviewed 07a7ee4fa097210a43e2ca9ce7f68b659a2cedb0 against stacked base 030e79fa6482acb08e51338ab4172ac4387b387d (producer/ms-marco-minilm-l6-v2-reranking). Current origin/main remains the charter baseline 0876e5ae1c98a169a6137e092e0d7b30bf9cee33; dependency PR #1322 remains OPEN/DRAFT at the exact expected head.

Blocking owner actions

1. Tester; Planner if scope revision is needed: CPU/fp16 is host-reachable, but L1 has no concrete latency, throughput, or memory result. The 2,700-second run and bounded 600-second retry are valid TIMEOUT-at-scale evidence, but they do not satisfy the required reachable precision performance tuple. Run a bounded reduced-iteration probe that emits at least one sample and memory result. If that cannot complete, Planner must explicitly revise the required precision/goal and invalidate downstream coverage claims; do not classify the reachable tuple as deferred.

2. Producer for dependency PR Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322; Tester: the shipped default reranking evaluator path is not operable. Independent command winml eval -m <reviewer-fp32>/model.onnx --model-id cross-encoder/ms-marco-MiniLM-L12-v2 --task reranking --ep cpu --device cpu --samples 2 loaded the 55,578-row pinned default dataset, then exited 1 because the dataset has relevant passage IDs but no candidates_column/materialized candidate text. The accepted L3 harness used direct ORT plus RerankingMetric; it did not exercise WinMLRerankingEvaluator. Make the default/public path materialize or consume bounded candidates, add a real CLI regression, and rerun bounded final-SHA winml eval through the shipped evaluator.

3. Tester, Learner, Explainer: recompute coverage after items 1–2. The current partial claim has zero deferred HOST-BLOCKED tuples; its only unresolved tuple is reachable CPU/fp16 L1 TIMEOUT-at-scale, which does not match the contract's full/partial computation. Update the verdict table, any changed findings, and the public body/reproduce commands; preserve DRAFT.

Verified gates

• Conversation: 0 issue comments, 0 line comments, 0 review threads, 0 open threads.
• GitHub checks: PR recipe(ms-marco-MiniLM-L12-v2): add CPU reranking recipes #1334 has 1/1 visible check successful (license/cla); dependency Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322 has 9/9 visible checks successful. Both heads were refetched after local review and did not move.
• Stack scope: exactly two files: examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/reranking_fp32_config.json and reranking_fp16_config.json. README/source/tests unchanged; both recipes are byte-identical to L6 prior art.
• Independent reruns: schema PASS; Optimum probe PASS; fresh fp32 recipe build PASS in 75.3s; ONNX structure PASS with named [1,512] int32 inputs and [1,1] float logits; license PASS; Ruff PASS; mypy PASS (438 files); models 1,534 passed; commands 3,638 passed; remaining 871 passed.
• Artifact review: tester seals have zero mismatches; no random-head warning; fp16 has 204 FLOAT16 initializers and is 66,856,699 bytes versus fp32 133,589,610 bytes, with fp32 I/O preserved. Named-input L2 parity passes both precisions; the bounded direct-ORT fp32 L3 smoke semantics are valid.
• Body/knowledge: required hierarchy and seven evidence subsections are present, commands are portable, and no scratch path appears. Lane A manual path is acceptable for this no-methodology-edit run: pushed commit e4ae80a75cf91bdc2c7f52336c46d3a00aa085a0, compare URL, and body file are present.

Reachable vs deferred

• Reachable verified: CPU/fp32 L0/L1/L2 and representative direct-ORT L3; CPU/fp16 L0/L2.
• Deferred: none.
• Reachable unresolved: CPU/fp16 L1 TIMEOUT-at-scale with no sample statistics.

[ssss141414]

Addressed the three requested changes on candidate 8dae90fbd75c30c79d0bd5ec60f7817d23d5bd6d, stacked on repaired dependency PR #1322 at 3708969b731425b0c6d4b97920d1b5e6519bb013:

• R1: CPU/fp16 L1 now PASSes a bounded true-fp16 CPUExecutionProvider inference: 187.936900 ms, 5.320934844 samples/s, RSS 638390272 bytes before and 686682112 bytes after (+48291840 bytes). Peak RSS was not measured or claimed.
• R2: the public winml eval path now reaches WinMLRerankingEvaluator on pinned streaming mteb/scidocs-reranking revision 56a6d0140cf6356659e2a7c1413286a774468d44: 2 groups/20 pairs, MRR@10=1.0, Recall@1=1.0, Recall@10=1.0. Dependency Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322 owns the generic default/schema/materialization repair; recipe(ms-marco-MiniLM-L12-v2): add CPU reranking recipes #1334 remains exactly two recipes.
• R3: coverage is recomputed as full, with both required CPU fp32/fp16 tuples passing and deferred/unresolved tuples empty. The PR body now uses tester verdict v2, learner findings v2, the updated dependency/current-main provenance, and the exact portable public reproduce commands.

The PR remains draft with model-scale-by-skill; title, base, label, and two-file scope are unchanged. This reply records the fixes for reviewer reassessment and does not claim approval.

[ssss141414]

REQUEST_CHANGES

Reviewed head 8dae90fbd75c30c79d0bd5ec60f7817d23d5bd6d against dynamic base 3708969b731425b0c6d4b97920d1b5e6519bb013 (PR #1322); current main/charter baseline is 0876e5ae1c98a169a6137e092e0d7b30bf9cee33.

Blocking owner action

1. Tester - regenerate the sealed fp16 evidence root with a truthful terminal phase. The sealed fp16-one-run.json says PASS, the outer stage records exit 0, and independent review reproduced the inference, but the same sealed root's fp16-phase.json says {"phase":"failed","error":"SystemExit(0)"}. The controller wraps raise SystemExit(main()) in except BaseException, so it overwrites its own complete marker after successful execution. Create a new revision-specific immutable tester root, correct the controller finalization (do not mutate the sealed root), rerun the one-iteration fp16 probe, reseal it, and regenerate the tester/downstream handoffs. Update the PR body only if public values change.

Re-entry verification

• R1 behavior closed: reviewer-owned true-fp16 CPU run used artifact SHA-256 d0bd17a024cca4ebf7b36dae607e87031d1cdff9d31d52310632825154c13ae6, 204 FLOAT16 initializers, named input_ids/attention_mask/token_type_ids, 0 warmups, exactly 1 session.run, and an 1800-second cap. Result: 215.447100 ms, 4.641511 samples/s, RSS 638717952 -> 687263744 bytes (+48545792), scalar 8.13815689086914, CPUExecutionProvider.
• R2 closed: line-by-line review of dependency repair 030e79fa..3708969b found generic pinned streaming SciDocs defaults, grouped positive/negative schema detection, positive-first bounded materialization, preserved pairwise/materialized behavior, and no model hardcode. Reviewer public CLI Eval reached winml.modelkit.eval.reranking_evaluator.WinMLRerankingEvaluator: 2 streamed groups, 20 pairs (5 positive + 5 negative each), MRR@10 1.0, Recall@1 1.0, Recall@10 1.0.
• R3/body closed: required hierarchy and all seven validation subsections are present; exact head/base/main and Lane A d86f80191a4e0ff713885e7408020b5c82b3c22f are reported; commands are portable; prior timeout/direct-ORT claims are clearly historical; coverage derives as full, with deferred [] and unresolved [].
• Coverage split: CPU/fp32 and CPU/fp16 pass L0/L1/L2; one representative CPU/fp32 public evaluator smoke passes L3. No deferred tuples.
• Scope/artifacts: exactly two recipe additions; README/source/tests unchanged; both recipe blobs are byte-identical to 07a7ee4f; all 87 tester seal entries hash correctly; build/analyze/HTP/config artifacts were read directly.
• Independent quality reruns: license PASS; focused reranking 109 passed; mypy clean in 438 files; commands/config/build/compiler/session/eval 3641 passed, 9 skipped, 1 warning; datasets 78 passed; utils 279 passed, 1 skipped. Lock-equivalent Ruff PASS. Locked Ruff 0.15.21 also reports five RUF036 diagnostics, reproduced identically on clean current main in untouched _live_chart.py/inference/engine.py, so they are pre-existing baseline drift rather than this stack's regression.
• GitHub final gate: recipe(ms-marco-MiniLM-L12-v2): add CPU reranking recipes #1334 has 1/1 visible check successful; Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322 has 9/9 successful. Heads and main were refetched unchanged after local review. PR remains OPEN/DRAFT with model-scale-by-skill.
• Conversation gate: 2 issue comments, 0 line comments, 0 reviews, 0 review threads, 0 open threads. Prior request 5382341341 and explainer reply 5383092316 were both re-enumerated; the reply substantively addresses R1-R3.
• Process audit: 0 reviewer-owned processes remain.

This is an internal evidence-integrity correction only; no product-code or recipe defect was found, and the PR must remain draft.

[ssss141414]

R4 addressed in a fresh immutable Tester root: the controller was corrected, the bounded true-fp16 CPU inference reran at the unchanged 8dae90fb candidate, and terminal state is consistently PASS / completed / exit 0. Cross-file validation and the 26-entry seal have zero mismatches; the old sealed root remains preserved only as incident provenance. Tester v3 is authoritative, and Lane A separately records _meta-111 in pushed commits 2c5700a047fc85e5fb3405d361f8fb06e1c846fd and e70be83aefd621dbb4170aa647fe6344a6d4884c (manual compare: https://github.com/gim-home/ModelKitArtifacts/compare/main...ssss141414/learner-l12-bert-knowledge-20260823?expand=1).

[ssss141414]

APPROVE

Reviewed head 8dae90fbd75c30c79d0bd5ec60f7817d23d5bd6d against dynamic base / PR #1322 head 3708969b731425b0c6d4b97920d1b5e6519bb013; current main is 0876e5ae1c98a169a6137e092e0d7b30bf9cee33. The final re-fetch was stable and the PR remains OPEN/DRAFT with model-scale-by-skill. Leave it draft.

Verified outcome

• Full coverage. Both required tuples, CPUExecutionProvider/cpu/fp32 and CPUExecutionProvider/cpu/fp16, are fully verified at L0, L1, and L2. The required representative L3 functional smoke is PASS on CPUExecutionProvider/cpu/fp32. Deferred []; unresolved [].
• R1 closed: true-fp16 artifact SHA-256 d0bd17a024cca4ebf7b36dae607e87031d1cdff9d31d52310632825154c13ae6, 204 FLOAT16 initializers, one bounded CPU inference, 187.938499962911 ms, 5.320889547364412 samples/s, RSS 647761920 -> 695275520 bytes (+47513600), no peak-RSS claim.
• R2 closed: the unchanged dependency head provides public WinMLRerankingEvaluator SciDocs support and its tests; independent same-head CLI evidence covers 2 groups / 20 bounded pairs with MRR@10 1.0, Recall@1 1.0, Recall@10 1.0.
• R3 closed: coverage recomputes to full, with no deferred or unresolved tuple.
• R4 closed under _meta-111: authoritative corrected root records verdict/result PASS, phase completed, outer exit 0, timed_out: false, and no error. Candidate/base/artifact/provider/root identities agree; all referenced paths exist; independent rehash found 26 manifest entries, zero mismatches, manifest SHA-256 8a72da5c7f308ffae9bcd60fcf09b83b0ad5bf339a44dd5f6c40512e544e72ca, and post-seal PASS. The old contradictory root is preserved only as incident provenance and is absent from authoritative required paths. A reviewer-owned success/failure harness self-test also passed.

Gates

• Delta is exactly the two L12 CPU recipe files; production recipe README, source, and tests are unchanged in PR recipe(ms-marco-MiniLM-L12-v2): add CPU reranking recipes #1334. Recipe blobs and artifact hash match the prior independent same-SHA review.
• Exact-SHA local gates remain applicable: license PASS; mypy PASS across 438 source files; focused reranking 109 passed; commands 3641 passed / 9 skipped; datasets 78 passed; utils 279 passed / 1 skipped; no introduced Ruff diagnostic (the five locked RUF036 findings are identical on current main).
• Final CI: PR recipe(ms-marco-MiniLM-L12-v2): add CPU reranking recipes #1334 1/1 COMPLETED/SUCCESS; dependency PR Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322 9/9 COMPLETED/SUCCESS.
• Discussion enumeration with pagination: 4 conversation comments before this verdict, 0 line comments, 0 reviews, 0 review threads, 0 open threads. Both prior requests (5382341341, 5383245876) and fix replies (5383092316, 5383380028) were verified.
• Lane A _meta-111 methodology shipment is remotely verified at final e70be83aefd621dbb4170aa647fe6344a6d4884c, with implementation 2c5700a047fc85e5fb3405d361f8fb06e1c846fd; manual compare path accepted.

Blockers: none. This is a skill-level opinion comment only; it does not alter GitHub review state or authorize readiness changes.