feat(groundedness): retrieval-quality scorer

[drewstone]

What

Lifts groundedness — retrieval-quality scoring — into the substrate (./groundedness subpath).

Did the retrieval/search provider actually return the fact the task needed, independent of whether the agent then used it. Isolates provider quality from agent skill (a perfect result + weak model still fails; junk results + strong model recovers — so task pass-rate alone conflates the two).

• Pure scorer scoreGroundedness(resultText, requiredKnowledge[]) → { score, found, missing, total, hadResults }.
• Trace extractor over the substrate's own retrieval/tool spans (RetrievalSpan.hits[].content + provider ToolSpan.result) — the provider tool matcher is injectable (defaults to search/research-not-fetch).
• Structural sibling of authenticity (pure deterministic scorer + consumer-supplied domain config); subpath-only export, no root re-export.

Why

A fundamental retrieval-eval primitive every retrieval-augmented eval needs; the provider's job is to surface the answer, this measures whether it did. Distinct from output-coverage scoring (which scores the agent's produced artifact) — this scores the provider's retrieved text.

Tests / checks

tsc --noEmit --strict 0 errors, vitest 10/10, tsup emits dist/groundedness.

[tangletools]

✅ Auto-approved PR — 11677e14

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-24T12:33:15Z}