[ProfSynapse]

Stacks on #266 (feat/native-mlx-execution-r2), which stacks on #265. This branch includes base de22015ea77180679f65d090e88060072a4ce040; no main merge.

Scope

• Compose the existing public TrainingAPI/RunsAPI, Foundation, reference host ports and SQLite store under examples/training_qualification. One restricted initial attempt requires an external root signature over the exact compiled workload, host selection and private state. Ordinary native qualification remains closed.
• Include the four accepted PR266 repairs: clean pinned-base saved-final adapter reload with exact tensor comparison; preclaim runtime/model/tokenizer revalidation; exact LoRA target and layer artifact layout; complete tokenizer selection and local loadability from published bytes.
• Include the accepted PR269 fixes: signed compiled workload/full host selection, authenticated cold reads after dispatch expiry and remote ref advance, and bounded owned host subprocess draining including interruption cleanup.
• Add a reusable private model preparation candidate at examples/training_qualification/native_model_preparation.py: local inspect, a bounded SDK metadata candidate, and later bounded exact-member acquisition into create-only model/tokenizer snapshots. Its receipt says OBSERVED_NOT_ADMITTED and grants no training authority. No model SDK action was executed in this round.

R5 Mac preparation and evidence

The accepted prior head aba3c57699f6617878440b3126d15e475e295651 was cloned as a clean isolated pushed source into a fresh 0700 Mac stage. The original private CPython 3.11.14 candidate and 50-dependency lock were inspected read-only; uv pip check passes. The M4 Mac reports 24 GiB and macOS 27.0.1. Two genuine synthetic publications were built with the existing ingest/prepare-dataset CLI and copied to private Mac staging, each with 3 training and 2 validation rows. Publication digests are 0e4951bb2e54ed0985e6ca0d4c9940f0f5d75ccb78821c14544ef44897580cbe (raw) and d22b95afbf5e415fa6a363381fda7714cfd0d7817d3b98a5e44e3149250ee9a0 (messages). Full details and commands are in examples/training_qualification/native-host.md.

The static capture correctly refused the group-writable /opt/homebrew/Cellar ancestor, without producing a capture. The candidate venv also lacks jsonschema required by the current host composition. Cached unpacked wheel metadata does not attest the absent original wheel archives. The model workspace remains empty and its acquisition draft has no selected members. Exact tokenizer IDs/assistant labels, selected install archive provenance and current-source capture remain unproved. The new head must be staged and recaptured after root review.

Verification

The earlier PR269 runtime/host repairs were accepted independently at aba3c576. At new head 8baceba2f08c4750ecc4c96a3f3aae20f0c58078, 29 focused model-preparation and host tests pass; Ruff, syntax and skill mirror checks pass. No locked runtime member, dependency pin or inventory changed, so the prior seven closure checks remain the applicable source baseline. The new preparation helper and Mac evidence need root review before any further effect.

No MLX import, model weights download, model load, training, GPU operation or provider/cloud call occurred in R5. No signed execution admission exists. Public logs/cancel/reconcile and GRPO remain unsupported.