Backfill explicit discovery metadata on sample YAMLs; thin the derivation heuristics

[staging-devin-ai-integration]

Background

PR #544 introduced presentation-only discovery metadata (group/variant/category/tags) for sample pipelines. These are first-class optional YAML fields, but in practice almost every sample (≈85/88) relies on the derivation fallback in apps/skit/src/sample_discovery.rs, which infers grouping/variants/category/tags from:

• hardcoded filename token tables (COMPOUND_TOKENS, SINGLE_TOKENS, LANGUAGE_TOKENS) in Rust, and
• a parallel SYNONYM_GROUPS table in TS (ui/src/utils/samplePipelineOrdering.ts).

This re-encodes codec/hardware/language knowledge that already exists structurally in each YAML's node kinds, and splits it across two languages.

Problem

A new codec, hardware backend, or language silently mis-groups until both token tables (Rust + TS) are edited. The heuristics are doing the work of the default path rather than acting as a thin safety net.

Proposal

Backfill explicit group/variant/category/tags on the sample YAMLs (a co-located, reviewable change), then shrink the derivation in sample_discovery.rs to a minimal fallback for samples that omit them.

Why this was deferred from #544

S4 was explicitly scoped to need minimal/zero edits to sample YAML files to avoid merge conflicts with the parallel sample-testing work (merge order S2 → S3 → S4). Bulk-editing ~85 samples during #544 would have collided with those branches. Now that they've merged, this backfill is safe to do as a focused follow-up.

Source: Devin Review observation on PR #544 (sample_discovery.rs "altitude" finding).

[staging-devin-ai-integration]

Additional findings folded into this issue (PR #544 second validation pass)

Two further observations from PR #544 are best solved by this backfill rather than more UI/heuristic patching:

1. Non-canonical card title (ui/src/utils/samplePipelineOrdering.ts groupSamplePipelinesByScenario)

A scenario group with no variant-less (canonical) member titles its card with an arbitrary collation-first variant: base = variants.find(v => !v.variant) ?? variants[0]. The pill-label fallback (baseVariantLabel) was fixed in #544, but the card header still guesses.

Shipping example: the dynamic speech-translate family (en-es, es-en, helsinki-en-es, helsinki-es-en) all derive a variant, so none is canonical. The collapsed card titles itself "Real-Time Speech Translation (English → Spanish)" (first sorted variant) and then repeats that same variant as one of four pills. Cosmetic; selection works.

Cleanest fix is server-side: emit a canonical-member marker or a neutral group label so the UI doesn't guess — which is exactly what explicit group/variant metadata on the YAMLs provides.

2. is_translation rides the same substring heuristic (sample_discovery.rs)

Translation grouping, the directional variant label, and the Translation category all now hinge on the derived translation tag, which tags_for_kind sets only by substring-matching nllb/helsinki node kinds. A future translator published under a different kind (e.g. plugin::native::m2m100) silently gets is_translation = false → language tokens aren't stripped → speech-translate-en-es and -es-en derive distinct group keys and don't collapse, lose directional labels, and get no Translation category — with no failing test.

Same hardcoded-substring fragility this issue already covers, now extended to the is_translation switch. The durable fix is a single server-side capability source (ideally registry-attached) feeding tags + human labels, plus explicit YAML metadata as the authoritative grouping signal.

Source: Devin Review second validation pass on PR #544.

[staging-devin-ai-integration]

The explicit-contract rewrite this issue tracked has landed in #544 (commit 3db13ca), so the heuristic-derivation path is gone rather than thinned:

• Runtime derivation removed from apps/skit/src/sample_discovery.rs — no filename tokenization, no substring category/tag inference, no codec sniffing, no canonical guessing.
• Sample YAML is now the source of truth: every bundled dynamic/ + oneshot/ sample authors category/tags and, where grouped, group/variant/canonical plus optional search-only keywords.
• The server emits the authored fields plus a resolved search_terms document; the UI does plain substring matching and the TS SYNONYM_GROUPS table is deleted.
• Card identity comes from the group's canonical member, fixing the non-canonical card-title issue (previously variants[0]).
• is_translation substring fragility is moot — translation grouping/category/variant are authored, not inferred from nllb/helsinki node kinds.
• A new build-time test (apps/skit/tests/sample_discovery_metadata_test.rs) fails CI if a bundled sample is missing required metadata or a group is internally inconsistent.

Closing once #544 merges.