feat(cli): sync ingest schema; wire sentence_pair_classification + embeddings

[saadqbal]

Summary

Clears the failing Schema drift check that's red on every open PR (including #160). internal/schema/ingest.v1.json had drifted from upstream tracebloc/data-ingestors@master, which added two task categories (sentence_pair_classification, embeddings), a content_hashdata_id strategy, and description updates. The drift predates and is unrelated to any in-flight feature — it was stale on develop too.

What's in

• Re-vendor the current upstream schema via scripts/sync-schema.sh.
• Wire both new categories into categoryRegistry (internal/push/category.go) as FamilyText — both stage raw .txt from texts/, matching data-ingestors conventions.pyTEXT_CATEGORIES.
• Mark both CLISupported: false with an UnsupportedNote, mirroring how causal_language_modeling / seq2seq / token_classification were added — conservative until the tracebloc ingest discover/build staging path for them is confirmed.
• Update the pinned lists in category_registry_test.go.

⚠️ Reviewer decision

I defaulted both to not-yet-CLI-supported. If dataset push already stages the text_a\ttext_b (sentence-pair) and anchor\tpositive[\tnegative] (embeddings) texts/ layouts, flip CLISupported to true and drop the note.

Testing

scripts/sync-schema.sh --check clean; go build/vet/test ./... green. Fixes TestRegistryCoversSchemaCategories, which a bare schema sync would otherwise break.

Closes#162

🤖 Generated with Claude Code

---

Note

Low Risk
Mostly schema vendoring and registry metadata; no new push staging paths or runtime content_hash logic in this diff.

Overview
Fixes schema drift by updating the vendored ingest.v1.json to match upstream data-ingestors: new categories sentence_pair_classification and embeddings, a content_hash option for data_id.strategy, and expanded texts / validation rules (supervised sentence-pair requires label; embeddings is self-supervised and must not set label).

The CLI categoryRegistry gains both categories as FamilyText with CLISupported: false and UnsupportedNote text (same pattern as CLM/seq2seq), so schema-valid categories are recognized instead of rejected while ingest discover/build for their texts/ layouts remains pending. category_registry_test.go pins the expanded known set and unsupported list so TestRegistryCoversSchemaCategories stays green.

^{Reviewed by Cursor Bugbot for commit d7574c9. Bugbot is set up for automated code reviews on this repo. Configure here.}