[WIP takeover] Preserve three-commit local runtime patch history

[robotlearning123]

WIP takeover branch — do not merge together with takeover/cosmos-runtime-dirty-wip-20260721. Both branches have the same final tree; this branch preserves the existing three-commit salvage history.

Commit structure

1. 2690f8113e13e91c92c1ee79010db079fed64eb4 — honor an explicit tokenizer cache directory
2. 5e6ec3761c992bda00f592a7a3e8d9cd048981bc — support cached text embeddings in the local video dataset
3. 1ed6206b27d5257f0b39eacd88e4f563f29b570f — crop mixed-length visualization batches

The source worktree remains clean on salvage/local-runtime-patches-20260721; this takeover ref points at the same head without moving that branch.

Validation

• PASS: AST parsing of all four changed Python files
• PASS: Ruff check and format-check on all four files
• PASS: gitleaks scan of the four-file diff
• PASS: largest changed file is 18,052 bytes; no cache or ordinary-Git oversize file is included
• BLOCKED: no focused unit tests exist for these four paths
• BLOCKED: the local pytest harness rejects configured option --suppress-no-test-exit-code, indicating the matching pytest plugin/environment is absent

Handoff

This is the preferred review history because each concern is separated. The branch remains WIP until targeted dataset, tokenizer, and callback tests are added and run in the intended Cosmos dependency/GPU environment.