Regenerate dataset card from current recipe

Update recipe: docs/corpus_blend.md

Sync train/corpus.py with 2026-08-14 document-separator fix (new blend sha256)

Sync scripts/prepare_corpus.py with 2026-08-14 document-separator fix (new blend sha256)

Sync scripts/blend_corpus.py with 2026-08-14 document-separator fix (new blend sha256)

Sync docs/measurements/blend_manifest.json with 2026-08-14 document-separator fix (new blend sha256)

Sync docs/corpus_blend.md with 2026-08-14 document-separator fix (new blend sha256)

Sync README.md with 2026-08-14 document-separator fix (new blend sha256)

docs: cross-link the now-public tt-tnt model and GitHub repo, fix stale (private) status

Correct dataset card: episod/tt-tnt is now trained on this nine-source blend, not TinyStories alone; record the tokenizer-pass token count

Add corpus recipe: source registry, reconstruction scripts, provenance docs

initial commit