Review-input truncation causes false 'missing screenshot evidence' closes

[JSONbored]

Advances #8828.

Confirmed production failure class, found during the 2026-07-26 decision-audit backfill adjudication.

The review pipeline truncates long PR descriptions before they reach the AI reviewers. When the truncation point falls before a PR's Screenshots section, reviewers conclude the required visual evidence is absent and the gate closes the PR for a policy violation the contributor did not commit.

Evidence:

• feat(ui): add an RSS feed discovery link to the site footer (#3351) metagraphed#4640 — closed at confidence 0.5 for missing before/after screenshots; the PR body verifiably contains 6 image attachments. The finding even says the description "is truncated exactly at that section." Adjudicated incorrect / policy_misapplied in the audit ledger.
• feat(ui): add network decentralization scorecard to /status metagraphed#3760 — same shape: the finding admits the description is "truncated immediately before what appears to be a Screenshots section." Adjudicated uncertain.

Requirements (model/fact boundary — same doctrine as #8833):

1. Whether a PR body contains image/video attachments is a deterministic FACT, not a judgment call: extract an attachment_count/has_images fact from the FULL untruncated body (cheap regex over ![](...), user-attachments, image extensions) and inject it as a structured fact the prompt cannot contradict — reviewers must never infer evidence-absence from prose.
2. When the description IS truncated, any evidence-absence finding (screenshots, linked-issue text, checklists) must be demoted to a nit/question, never a blocker — absence of evidence inside a truncated window is not evidence of absence.
3. Regression tests: a long-body fixture with images beyond the truncation point must produce zero evidence-absence blockers.

Test coverage: 99%+ branch-counted on changed src/** lines, both truncated/untruncated arms and with/without attachments.

[linear-code]

JSO-6208