[david6666666]

Superseded by the requested split: #110 — h3.c repository knowledge (merged first) and #111 — generic research framework (second). Neither replacement includes the excluded plan document. This combined PR is closed without merging.

---

Governing Issue and delivery slice

Refs #103

knowledge search ANDs all query tokens, so concatenating a model, framework, diffusion and video can hide transferable knowledge. This adds opt-in knowledge research: separate model/framework/task/modality/mechanism queries, a bounded deduplicated union, and per-result query attribution. Existing exact search and package filters retain their contracts.

The requesting owner explicitly authorized these code/workflow changes and a PR. This slice also updates the canonical engine-research procedure with eligible native-reference selection, complete component accounting, compatible quantization/engine pairs and conditional hardware decisions. It does not implement an H3 engine or Metal runtime, download model payloads, or claim target performance/quality acceptance.

Knowledge authorship declaration

{
  "schema_version": 1,
  "head_sha": "f804ec1503e977787c572d5d27d1f3258017b01f",
  "human_only": false,
  "material_ai_authors": [
    {
      "tool": "Codex",
      "model_id": "gpt-6.1-sol",
      "effort": "max",
      "role": "author"
    }
  ]
}

Changed artifacts and evidence

Requirement	Delivered artifact	Evidence and limits
Search model plus task/modality/framework knowledge	research.py, public CLI, packaged H3 profile	Independent queries retrieve H3, MLX and transferable diffusion/media knowledge; filters and provenance remain intact. Unknown models need explicit source-grounded dimensions.
Select a native implementation reference before engine generation	Canonical research skill and pinned H3 worked example	Inspected three eligible engines; h3.c has the highest observed star count within that set. The example uses direct source inspection; repo-to-skill is optional.
Account for all weights, quantization and hardware	Preparation procedure and pinned safetensors-header inventory	Selected official FL2VA files contain ~69.24B stored tensor elements including buffers and 144.02 GB of weights; selected MLX export differs. Documents full/truncated language/vision encoder, DiT and both VAEs, storage versus phase peaks, format compatibility and reported quality/hardware limits. Header-only metadata inspection does not establish payload integrity or M5/64 GiB feasibility.

Checks and dependencies

Tested content: f804ec1503e977787c572d5d27d1f3258017b01f; assessed base: 9e20eb7226de49f7368c4a24599f9ccf11ccda3c.

• Passed: 54 focused research/catalog/skill/contract tests; ruff check src tests tools; router skill validation; git diff --check.
• Passed: isolated wheel build, clean noneditable wheel[skills] installation, tests/check_installed.py with -I outside the checkout, and the actual installed console command. The profile, procedure, worked example and inventory are present.
• Windows full suite with UTF-8, an explicit source import path and a short temporary directory: 591 passed, 1 failed, 14 skipped, 70 subtests passed. The remaining test_unreadable_process_table_fails_before_launch failure mocks Linux on Windows, where os.getpgrp is absent; the implementer reproduced the same failure at the unchanged base. The test and process implementation are unchanged. This is not a local full-suite pass. The initial unconfigured run was 580 passed / 12 failed / 14 skipped.
• GitHub check evidence for this head: draft lint/smoke, Python 3.10/3.11 full checks and isolated installations, and knowledge declaration policy. Current check status is shown on this PR.
• Independent read-only review of the exact head/base found no actionable findings. The reviewer independently passed all 13 research tests, checked retrieval/filter compatibility and resource coverage, and verified selected remote header hashes/counts/bytes, source pins, inventory arithmetic and license notices. They did not rerun the complete suite or validate Apple hardware.
• No Mac inference, measured target peak memory, joint audio/video quality or end-to-end performance validation was run. Source reports remain unverified on the target.
• Existing registered skill inputs and generated abstracts are unchanged. Related backend/runtime work remains separate from this preparation slice.

Remaining work and Issue closure

The preparation slice is implemented and independently reviewed; current CI must be checked before readiness. Mac execution, media quality and target-memory qualification remain work for a separately frozen campaign. Goal acceptance and Issue closure remain with the requesting owner/maintainers; this PR uses a non-closing reference and does not claim campaign PASS.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	🔄 Running since 2026-10-23T02:50:17.609288Z	f804ec1	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.