Search- and selection-driven curation (curate --ids/--from, search --save)

[arpitg1304]

What

Closes the seam between semantic search / visual review and curation. forge curate was SQL---where only, so you couldn't curate "the top-K of a search" or a hand-picked set from Studio — those are selections (ranked lists / picked sets), not predicates. Now the selection can also be an explicit --ids list or a --from <selection.json> file.

# search → save a selection → curate it (provenance recorded)
forge search "regrasp after a failed pick" -c ./cat --top 20 --save sel.json
forge curate -c ./cat --from sel.json --label approved

# or an explicit id list
forge curate -c ./cat --ids e1,e2,e3 --label approved

Changes

• curate() gains an ids selector — composes with --where (intersection) and the dedup policy, and is filtered to episodes that actually exist, so a typo/stale id never creates a dangling label.
• forge search --save sel.json writes the result ids + provenance ({episode_ids, source}); forge curate --from sel.json applies them and records the source in labeled_by for auditability.
• Forge Studio's dedup tab gains a "Download decisions" button that exports the same selection file (the rejected picks) — so keep/reject review can write back via forge curate --from, instead of being copy-only.

Why this shape

The two real requests ("top-K of a search", "what I picked in Studio") are explicit selections, so an id/selection-file path fits them natively and unblocks both in one small change. A SQL similarity predicate (the alternative) is a filter, not a ranker, still needs the model outside SQL for text, and does nothing for the hand-pick case — so it's a bigger lift that solves less. Provenance on the label (labeled_by = the search/source) keeps curation auditable despite not being a self-documenting SQL rule.

Doing this before Phase 4 (snapshots) matters: snapshots filter on curation_labels regardless of how they were produced, so search/Studio-curated sets become exportable too, not just SQL-predicate ones.

Testing

5 new tests (curate by ids, id-filtering to existing episodes, ids ∩ where, and the search --save → curate --from round-trip with provenance); Studio template JS re-validated. No regressions.

🤖 Generated with Claude Code