Admit verified local SmolVLA datasets through normal preflight

[rylinjames]

Local corrected-dataset artifacts could not reach Tether's normal training preflight: the CLI accepted only an HF dataset ID/revision, and metadata checks always used Hub reads. This adds explicit --dataset-root and --dataset-manifest-sha256 inputs for one bounded lane: pretrained SmolVLA LoRA using a completed verified-moments-v1 export.

A stdlib verifier binds the canonical completion receipt, complete file inventory and actual hashes/sizes, v3 feature/statistics shapes, selected development split/provenance, exact declared producer pins and explicit SmolVLA normalization evidence. Missing local files and malformed receipts fail before metadata lookup; they never silently substitute Hub dataset metadata during Tether admission. Existing schema/padding and episode-count checks read local info.json, verify real selected-base metadata, and remain blocking for local input. Command construction revalidates before adding dataset.root.

Local requests reject skip-preflight, resume, revision ambiguity, unsupported model/training lanes and extra-argument escapes that could change dataset selection, output ownership, normalization, features or LoRA method. Existing remote-only flows and positional FinetuneConfig callers remain compatible. ImageNet default behavior is preserved and reported separately from explicit SmolVLA VISUAL=IDENTITY.

SmolVLA is the first lane because it already has meaningful base/action compatibility preflight. ACT's from-scratch path lacks that contract and remains outside this change. Studio's companion draft resolves workspace-owned export IDs, verifies source/split/runtime identity and supplies a separate job-owned snapshot; this public verifier alone is not an authenticity signature or filesystem race-proofing mechanism. Downstream LeRobot retains its own fallback behavior if callers later mutate/remove files, so the input must remain immutable through use.

Validation: 172 focused tests passed, including 111 new admission tests and 61 existing finetune/preflight/resume checks. Ruff and compilation passed. Independent cross-repository review has no unresolved findings for this offline admission scope. New public fixtures deliberately use opaque synthetic Parquet slots and synthetic base metadata; they do not establish loader or training qualification. Companion Studio verification in FastCrest/tether-studio#134 uses a real exported SmolVLA artifact and passed 23 offline checks plus one targeted recovery regression, stopping before trainer dispatch.

The original top-level tether.cli registration test also passed separately after its missing exact-candidate source/import closure was materialized and all 19 required Git blobs were verified. That check made zero network/subprocess attempts or model-runtime imports; no code or dependency edits were needed. The 172 tests were reused. The edited finetune Typer command also passed a dry-run with both new options. Full package/doctor and training acceptance are not claimed. GitHub Actions were not inspected or used.

Base main: 18157ec. Reviewed composed candidate: 7606360 (tree 2d4cfa4fc01538a0858c0d97935658b8ef640cdc). It includes the exact CUDA evaluation fix from Studio pin 731779f as a second parent: local_runner passes verification_device="cuda" and the regression asserts it. Both files match that pin byte-for-byte. Seven affected offline evaluation tests passed with zero network/subprocess attempts or model/provider imports; the unchanged admission/CLI evidence above is reused. Independent review confirms the full tree preserves both source lines and the artifact-identity primitive remains byte-identical. Studio's runtime-pin update is a separate dependency step.

No model/checkpoint download, training, GPU/provider call, HF upload, credentials or production data. This source integration does not qualify actual training, platform packaging or customer acceptance.