feat(review): pure linked-issue satisfaction assessment builder (advisory finding)

[JSONbored]

Add the deterministic-shaped, AI-backed ANALYSIS core of #1961: given the already-fetched linked-issue text and the PR diff/title, produce a bounded advisory result {status: addressed|partial|unaddressed, rationale, confidence}. Pure function + prompt composition only — NO gate wiring, NO disposition change (that stays maintainer-only). Fail-safe: no issue text or model error yields no finding, and a low confidence never emits an 'unaddressed' call.

Deliverables

• New src/services/linked-issue-satisfaction.ts exporting an assessment builder that takes {issueText, prTitle, prBody, diff} and returns a public-safe {status, rationale, confidence} | null
• Confidence gate constant (mirror ai-review.ts AI_DEFAULT_CONFIDENCE handling in src/services/ai-review.ts:384) so an 'unaddressed' verdict below the floor degrades to no finding
• Fail-safe guards: empty/absent issue text, malformed model output, or thrown error all return null (never throw into the caller)
• Public-safe rationale filtering reusing the existing public-safe helper used by composeInlineFindings
• Unit tests covering every branch: each status, empty issue, low-confidence suppression, model-error path, both sides of the confidence floor (97% branch-counted)

References

• src/review/linked-issue-hard-rules.ts:1
• src/review/grounding-wire.ts
• src/services/ai-review.ts:384
• src/services/ai-review.ts:215

Part of #1961.

---

size: S · held for later release (gittensor:feature once activated)

[linear-code]

JSO-2273