pr created body - kimjooyoon/meta-ontology-go GitHub Wiki

Add body-context --model for source-owned record-field assembly. The same verified local loader as code generation selects the model's explicit input representation. Inspection records the source digest, model/weight identity, input contract and either READY_FOR_RANKING or the existing representation-decline reason. Prediction and candidate-test counters are explicitly zero; readiness describes representation compatibility.

The source still owns alternatives, types, stable IDs and finite cases. A compatible model ranks candidates during body-compose; the compiler evaluates them. Without a model, assembly keeps deterministic order. Existing non-model exports preserve their JSON shape. Other context kinds reject this model option and an explicit feature must match the loaded model.

• body-context, body-compose and help models provide runnable guides using the included own-model example and saved replay.
• The native release profile retains two Korean/English preflight reports and binds their model-input digest to actual construction and saved replay. Missing counters are rejected rather than inferred as zero. Original failed command bytes remain .failed-output.
• Scalar evidence becomes ten reports per platform, while the release corpus remains 26 cases. This feature follows the separately published 0.6.21 source; it does not rewrite that release.
• The usability principle was inspired by tinyjsapp's capabilities/requirements inspection: https://github.com/tarwin/tinyjsapp . No framework code, new model downloads or training are introduced. Existing own metadata and 446-byte packed weights are reused.

Clean candidate: 936c7d8e06ca162b0c490a831bfa23bc0a2bf9ad, Go 1.27.2, runtime SDK v0.2.26-experimental.

• TDD: model/help inspection initially failed; native profile initially failed to retain preflight reports. Regression implementations pass.
• Final affected race: CLI 136.121s; bodycodegen cached from the same unchanged code validated at 34.564s after clock-fix integration; release profile 7.914s. Affected vet passes.
• Clean native compiler: actual scalar release profile passes in 6.112s. Separate raw command observations retain two previews and eight Korean/English fixed/model construct/replay reports.
• Preview predictions/tests 0/0; actual model assembly calls 1, first fields 7/12, selected fields 12/12; other four native inputs 4/4. Saved runtime new calls 0. Fixed/model generated Go digest matches, and model input digest matches preflight, construction and replay.
• One field declines with THREE_FIELD_CHOICES_REQUIRED; mismatched features, corrupted weights, cancelled context and changed source/model/counters/field identity are rejected. The CLI guides execute successfully; models is the documented topic name.

These are recorded finite examples, with unknown training exposure. They establish input compatibility, case checks and replay behavior for these examples.

794 changed text lines across the API, CLI, native integration, regression tests and documentation, plus two small compressed original preflight fixtures. The soft 400-line cap and 40-nonblank-line cap for the connected export/native-smoke slots are exceeded to keep one source/model contract with its native and negative verification in a single reviewable change. Generated regions and the public 0.6.21 snapshot remain intact.