AI linked-issue satisfaction assessment

[JSONbored]

Phase 2 — Interactive control & issue intelligence

Context

We enforce deterministic linked-issue hard rules (src/review/linked-issue-hard-rules.ts) and fetch linked-issue text in grounding, but never assess whether the PR actually addresses the issue.

Goal

An AI advisory that answers "does this PR satisfy the linked issue's intent / acceptance criteria?"

Requirements

• Use the already-fetched linked-issue text; add a bounded AI assessment producing an advisory finding (addressed / partial / unaddressed + short rationale).
• Advisory by default; optionally gate-able via a new gate.linkedIssueSatisfaction mode (off/advisory/block, default off).
• Fail-safe (no issue / model error ⇒ no finding). Public-safe. Confidence-gated to avoid false "unaddressed" calls.

Deliverables

• Assessment path in the review pipeline + gate wiring + tests + docs.

Expected outcome

• Catches PRs that miss the issue's intent before merge; reinforces the issue-centric model.

Effort

M.

[linear-code]

JSO-2041