Immutable execution plans

For a serious run, freeze the resolved config and Parquet inputs before loading model weights:

miniverl plan --config verl-opd.yaml \
  --accept-local-reinterpretations \
  --out plan.json --offline
miniverl run --plan plan.json --offline

plan --out is still weight-free. It scans the actual Parquet source and atomically writes a deterministic JSON artifact containing:

• the source YAML byte digest and every ordered override;
• the full compatibility report and explicit reinterpretation acceptance;
• immutable student and teacher revisions;
• every Parquet file hash plus schema, content and row-count identities;
• the exact validated native RunConfig, loss semantics and physical recommendations;
• declared tokenizer revisions. Structural tokenizer identities remain declared_not_loaded until the bounded hardware probe loads them.

The plan's canonical digest excludes only its own digest fields. run --plan validates the schema and miniVERL minor-version boundary, recomputes that digest, rechecks the YAML and every Parquet byte, rescans the data manifest, and consumes the sealed native config. It never recompiles with new defaults. Model construction starts only after all checks pass.

The same plan digest is recorded in the run manifest, teacher-cache index, checkpoint state and checkpoint identity. A cache or checkpoint from another plan is refused. Direct run --config remains available for quick experiments, but --plan is mutually exclusive with config and override options.

Deliberate limits

An immutable plan does not download or hash model snapshots, allocate CUDA, or claim measured memory. Those identities become measured only through the explicit probe/materialization stages. Moving model revisions, unsupported verl semantics and unaccepted local reinterpretations prevent plan publication.