decider: add native CPU request compilation and response assembly

[Levius-Fubuki]

Purpose

Add the weight-free native CPU request compiler and response assembler for Decider-2B v11. This is the processing dependency of #110; it does not load tensor payloads, execute Qwen/CUDA, serve HTTP or advertise model readiness.

Choice supports the 255 released labels; Noul uses A/B labels and Score expands isolated levels. Ordered rendering/tokenization, calibration, usage and request limits preserve the pinned contract. Unsupported template/layout modes are rejected. This is the processing dependency of #110.

Integration and validation — 2026-10-09

Merged main 0d2521035107e35a2670b4df716f9728a074d520 into head e0f297ee61703392d91bc26595d5da59f7385589 with a normal merge. Retained both JEV-VL and Decider workspace members and qualified the Decider lock dependency as sha2 0.10.9; pinned package versions are preserved. Authored Decider source and test bodies are unchanged. Merge this processing parent before #110.

Fresh isolated-head checks pass: cargo fmt --all --check; cargo clippy --workspace --locked --all-targets -- -D warnings; cargo test --workspace --locked (134 passed, 0 failed, 20 ignored); cargo build --workspace --release --locked; strict MkDocs; CUDA Python helper regressions. No new GPU/reference/HTTP campaign ran. Main's ABI6/prefix integration is included but device execution was not verified on this macOS host.

Current main-based diff: 13 files, 1,046 authored source/test/build lines, 1,263 total changed lines. Counts are additions plus deletions; docs, fixtures, licenses and lockfiles are excluded from authored code.

Historical evidence

2026-10-07 sequential validation archive, asset decider-next-20261007-final-evidence.tar.gz, SHA256 9f57d969fc484c518d8ebac7b2dd38174e2c6954e7af04b9352550675053ae48. Contract source f59cca40462d5b63c3dbc909e6aefb147bd8f44e, main base 99865743d27316fbe81362dc0f0e6de6fda86284.

2026-10-08 raw verification evidence, SHA256 5df66bf3974d534875946168fba4d7fe543d08edbd658c3756f17d6f0c8f5773. Public/file config validation and two pinned contract opt-ins were verified at their recorded source revisions. These historical results do not establish CUDA integration parity at the current head.

Self-review

Agent-assisted full-diff review and independent conflict-resolution review found no outstanding issue. Contributor review does not replace maintainer approval.

• Reviewed the full diff and resolved findings.
• Checked CPU contract ownership, dependencies and focused scope.
• Ran applicable checks and stated unverified inference/performance.
• Checked description and claims against implementation and source-scoped evidence.

Exact-head GitHub checks

Head e0f297ee61703392d91bc26595d5da59f7385589: CI passes, Docs passes. Docs deploy is intentionally skipped for pull requests.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[hsliuustc0106]

Reviewed 5e9e5c0945333d2296fb2f3bc26544cd22928a19 against merge base 7f39ac40902c374803992407bb26eeba29c8a588. One reproduced P2 configuration-validation defect is attached inline.

Validation: workspace tests passed (82 passed, 0 failed, 10 ignored), formatting and strict all-target Clippy for omni-decider-native passed. The contradiction was reproduced through the production crate's public validator and the pinned upstream resolver at Mapika/decider@50d0be0d7cb43d2066965ce5fa7f3fe4e489a60f. No checkpoint or GPU execution was run. Existing workspace passes do not establish full request/tokenizer or inference parity.

[hsliuustc0106]

Re-reviewed at head ac8033611f (4 commits; merge base 99865743d). The P2 from my 2026-10-06 review — Config::from_value accepted layout="plain" together with chat_template=true, contradicting the pinned upstream resolve_layout — is fixed at this head: commit f59cca404 adds ("chat_template", false) to the mode-validation loop in src/models/decider/native/src/config.rs, and tests/decider/config.rs adds both requested regressions (public_validator_rejects_chat_metadata_on_plain_layout, file_loader_rejects_chat_metadata_on_plain_layout) plus the missing/disabled acceptance and non-boolean-type cases. Verified in the code at this head.

Beyond the fix, no actionable findings. The CPU contract crate (strict-JSON policy with duplicate-key/depth/range rejection, 255-label derivation with uniqueness check, segment-aware option tokenization, Score isolation, unique_tokens usage accounting, whole-request validation) matches the pinned-contract documentation, and the goldens are enforced by the registered opt-in test against tests/decider/data/{cpu,responses}.json. I also confirmed the crate sources here are byte-identical at the series tip (#113) except for lib.rs doc wording, so this review holds across the stack.

Readiness is complete: full template, all four self-review checkboxes checked by the author, and the stated counts (1,046 authored / 1,263 total) match my independent count. Artifact hygiene is fine (licensing files and consumed fixtures only).

CI (rust/build/benchmarks) passes on this head (observed). The author's local counts (94 passed / 13 ignored) and the pinned upstream resolver behavior were not independently reproduced — no Rust toolchain on this host — and no checkpoint or GPU execution was run. Existing workspace passes do not establish full request/tokenizer or inference parity; that evidence lives in the later series PRs.

Provenance: canonical .agents/skills/system1-omni-review/SKILL.md (SHA-256 58be3bc6…dde108) and its repository map read at trusted base 4a79980d8a75, plus CONTRIBUTING.md and docs/architecture.md. Remote head rechecked before posting.