4. Marking and correction limits

Conditional Text markers are explicit and non-printing, but can dirty the document and persist when saved. Group marking uses exact supported anchors, skips context-only findings and is capped at 1,000 planned ranges; native writes retain additional guards. Issue marks and Style Map replace each other and have their own clearing actions.

The solver can return no safe candidate. Eligible corrections are bounded by goal, freshness, document identity, supported property and side effects. Style reset changes one supported exact property; nested targets, ambiguous values and review-only evidence remain unavailable. A reviewed/addressed status never proves a text correction.

Eligible overset frame-fit alternatives can expand the terminal frame's bottom edge in a guarded live preview; typographic alternatives instead change the supported composition parameters. Queue for review restores a live preview before retaining a proposal, keeps one candidate per story and does not provide batch Apply.

The current source has a source-confirmed quick-control mismatch: Language & rules → More shows the middle Semantic strictness option as Balanced, while its handler accepts Normal. Use Settings → Typography policy → Normal and explicitly apply/rescan for that middle setting. No fresh native-host reproduction or plugin code fix is claimed by this documentation revision.

Recommendation boundaries

Recommendations use accepted inspection evidence. Only ordinary repeated spaces, exactly three dots and qualifying unambiguous dates can receive concrete replacement samples. Other findings receive contextual advice. Ambiguous or invalid dates, unknown locale and semantic-pattern-only findings do not receive guessed replacements. Stale advice requires rescan. The recommendation card introduces no Apply or batch correction action.