Tag input swallows commas into one mega-tag

Why Care?

Surfaced live on 2026-08-02 while curating the “Quantum Innovation & Computational Biology” thesis: the source biology-begins-to-tangle-with-quantum-computing ended up with the tag Quantum-Computing-Quantum-Innovations-Computational-Biology — three intended tags fused into one — plus a stray Quantum-Computing. Edit mode then rendered the fused value as one oversized chip. The operator had typed commas; the input ignored them.

Root cause

toDashed (apps/corpora-curator/src/curation.svelte.ts) is the culprit:

export function toDashed(s: string): string {
  return s.trim().split(/[^a-zA-Z0-9]+/).filter(Boolean).join('-');
}

It splits on any non-alphanumeric run — spaces and commas alike — then hyphen-joins everything back. So a whole comma-separated entry becomes a single dashed token.

Both tag entry points feed the entire raw input straight into it as one tag, with no comma-splitting:

• Source edit: TagBar.svelte → curation.applyTag(raw) → toDashed(raw) → one tag.apply.
• Corpus create: StrategyPicker.svelte addTag(t) → toDashed(t) → one pendingTags entry.

(The Authors field, by contrast, does comma-split — the inconsistency the operator reasonably expected the tag field to share.)

Expected behavior

Commas (and newlines) are tag separators; spaces within a segment stay word-joiners (dashed). So:

Quantum Computing, Computational Biology → ["Quantum-Computing", "Computational-Biology"]

Pasting a comma-separated list into either tag field should create multiple distinct chips, deduped, each Train-Case-dashed.

Fix

Add a splitTags(raw): string[] helper next to toDashed (split on [,\n]+, toDashed each segment, drop empties), and route both entry points through it:

• applyTag loops over splitTags(raw), firing one tag.apply per new tag.
• StrategyPicker.addTag appends every splitTags(t) result to pendingTags.

Suggestion-click paths pass a single token, which splitTags returns unchanged — no regression.

Resolution

Fixed 2026-08-02. Note the runtime fix only reaches augment.didi.sh after a corpora-curator rebuild + redeploy; the already-mangled data on existing records is corrected separately (the Quantum source’s tags were repaired directly in the canonical layer the same day).

See also

• [[feedback_tags_train_case]] — the Train-Case convention this preserves.