[riquelmechile]

Linked issue

Closes #89

PR type

• Bug fix (type:bug)
• New feature (type:feature)
• Documentation only (type:docs)
• Code refactoring (type:refactor)
• Maintenance / tooling (type:chore)
• Breaking change (type:breaking-change)

Summary

• Work unit 3 child W3C1 (auto-split, stacked-to-main on merged W3B2 test(business-domain): complete promotion authority fake coverage matrix #92, 396 ≤ 400 lines): the read-only app evaluation seam (task 2.1) + core tests (task 2.2 first half).
• packages/app/src/learning/evaluate.ts: activated tenant Skill → unique applicable policy → ONE event-stream read → aggregate → parse explicit+authority evidence → repository authority resolution for the trusted actor/principal → pure evaluatePromotion. Zero writes; empty tenant rejects before reads; policy failures stop BEFORE the event read; malformed input is typed unavailable; authority unavailability reaches needs-review via the evaluator, never promote.
• Deviation (documented in apply-progress + Engram): design listed deps:{events,skills}; the seam adds authority + trusted{principalId,actorId} because authority is repository-resolved, not caller-supplied.
• Strict TDD: RED (module missing) → GREEN 3/3 core (gold promote through the full seam; exactly one tenant-scoped read per repo + zero writes; policy-not-found stops before any event read); full gate pnpm check exit 0 — 1564 passed | 6 skipped, 0 new warnings.

Changes

File or area	Change
packages/app/src/learning/evaluate.ts	Read-only evaluation seam (new, 184)
packages/app/test/learning/evaluate.test.ts	3 core seam tests with recording doubles (new, 211)
openspec/changes/learning-promotion/tasks.md	Task 2.1 checked

Verification

• Focused checks pass: pnpm exec vitest run packages/app/test/learning/evaluate.test.ts → 3/3
• The affected runtime scenario was exercised, or is not applicable (pure seam over in-memory fakes; no PG in this slice)
• Full gate: pnpm check GREEN — 1564 passed | 6 skipped, 0 new warnings
• Native RDD: lineage review-6ad031598d906097 (medium, reliability lens) approved with focused+full-gate evidence; pre-pr gate allow

Contributor checklist

• The linked issue is open and has status:approved.
• Exactly one PR type is selected and its matching type:* label is applied.
• The branch name follows type/lowercase-description.
• Tests and documentation are updated where behavior changed.
• Commits use Conventional Commit format and contain no attribution trailers.
• No secrets or unrelated changes are included.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 414c0db2ab

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Return a candidate result when no outcomes match

When the selected window contains zero matching outcomes—a normal state for a newly activated skill—createLearningCandidate rejects its empty outcomes array, and this branch reports the evaluation as unavailable. That bypasses evaluatePromotion, whose threshold semantics return remain-candidate with insufficient-positive-observations; callers therefore cannot distinguish ordinary insufficient evidence from a malformed/unavailable evaluation. Handle the empty aggregate as an evaluated remain-candidate result rather than no-matching-outcomes.

Useful? React with 👍 / 👎.