[raiseCatError]

Long-command editing repeatedly constructs native grapheme segmenters and recalculates prompt width for each glyph. Visual-row movement additionally recalculates the full layout for every possible token boundary, producing a ~3-second arrow-key stall in the saved workload.

Reuse a lazily created segmenter, calculate prompt width once per row, and capture all caret positions in one pass through the shared wrapping implementation. Navigation retains token boundaries, paste-label mapping, earliest-token ties and clamping when clusters join across separately typed tokens. Temporary caret storage is linear; no input cache, dependency or worker is added.

Measured on M2 / 8 GiB / macOS 27 / Node 26.12, against unchanged dev:

• One visual-row moveUp in a long Unicode command: median 3,041.94 → 2.12 ms.
• 30 long-command edit/layout/highlight steps: median 201.97 → 46.75 ms.
• Built-frontend long-command typing-update p95: 45.86 → 8.78 ms; deletion-update p95: 13.06 → 7.01 ms (seven warm disposable sessions, 105 observations each). PTY observations, not physical key-to-pixel latency.
• None-prefix long-layout component median regressed 3.4% in this small sequential sample; no gain claimed there. No application idle/startup-memory gain claimed. Raw per-stage samples and conditions are in docs/performance/2026-10-25-input.md.

Deterministic resource/scaling regressions were observed failing before implementation, then passing. Compatibility tests enumerate token positions across Unicode, split combining clusters, wrapped paste labels, narrow widths and styled/empty prefixes. Independent read-only review approved with no concrete blocker. Focused input tests and build pass; full canonical npm run verify passed with explicit xterm-256color/truecolor capabilities, NO_COLOR unset, two workers and private short T7 fixtures. No test assertions/deadlines weakened.

Physical Ghostty/Zed QA: NOT RUN. Check long Unicode typing/deletion, wrapped/multiline vertical movement and selection, resized/styled/None prompts, multiple paste atoms and unwrap, Shift+Enter/Ctrl+J and Option+Backspace. Known intermittent macOS read issue is not claimed fixed. Keep the issue open pending hands-on validation.

Refs #507

Targets dev; do not merge automatically. Independent of output-rendering PR #506.

Exact-head CI: 10250f541f613153de10945aa11d3435cd92c2f4 passed macOS/Ubuntu shards, Quality and Node 22 compatibility; aggregate CI green. Fedora/timing jobs skipped by workflow conditions. CI run. PR remains unmerged.

Environment clarification: t7run supplies NODE_COMPILE_CACHE=/Volumes/T7Tmp/ai-dev/caches/node-compile; the original before/after runs inherited it. No OS-cache purge or true cold-start claim. Separate cache-disabled/private-cache startup controls are being retained in the local evidence.