perf: measure and reduce textarea document-wide work and paste scaling

[subinium]

Priority: P2
Evidence: Source-supported performance investigation; no release speedup measured
Baseline: v0.23.0 source 37933f9, audited 2026-09-05. Native execution/compile probes used Rust 1.98.0 on aarch64 macOS unless explicitly stated otherwise.

Problem and Reproduction

Each frame clones the line vector and builds a visual map; accepted paste repeatedly counts/indexes the growing line per grapheme. Undo retains bounded-count full-document snapshots. #95's no-second-wrap-on-unchanged-frame branch is present; the remaining costs are not the old double-wrap claim.

src/context/widgets_input/textarea_progress.rs:68 · src/context/widgets_input/textarea_progress.rs:317 · src/widgets/input.rs:1077

Scope

• Benchmark idle/edit/end-vs-middle paste on1KB/100KB/1MB ASCII/CJK/ZWJ documents.
• Consider bulk insert, versioned/lazy visual indexing and byte-aware/delta history only where justified.
• Preserve externally mutable text state, cursor/grapheme behavior and response.changed semantics.

Expected Effect

Reduce latency and memory copies for large editors while retaining exact editing and undo behavior.

Acceptance Criteria

• Publish release before/after p50/p95, allocations/bytes and history retained memory.
• No-op or net-zero batches do not spuriously report changes or evict history.
• Public line mutations and wrap-width changes invalidate caches correctly.
• Retain rejection/partial paste undo and mixed-event ordering regressions.
• Follow the repository quality gates and the evidence rules in tracking: post-v0.23 correctness, resource efficiency and WASM completion #366; record actual results rather than treating a proposed test as passed.

Boundaries and Dependencies

Do not substitute a naive dirty flag that changes net-result or public-mutation semantics, or mandate rope/arena adoption without evidence.

• input: invalidate textarea visual lines between edits and navigation in one batch #372
• input: preserve the original order of paste and key events #373
• input: do not evict undo history for a rejected textarea paste #374
• perf: establish release-mode workload, allocation and idle-resource baselines #414

Related closed work: #94, #95, #91. This is a scoped follow-up at the audit baseline, not a claim that every prior fix is absent.

Tracking: #366. This issue records unresolved work, not an implemented fix.

[subinium]

Scoped textarea changes and evidence are included in draft PR #433 and docs/PERFORMANCE.md. Bulk splice and borrowed unwrapped viewport work reduce 1 MB ASCII idle allocations from 24,549 to 54 in the measured fixture. Wrapped CJK idle did not improve in the recorded sample; full-document undo snapshots remain bounded by count, not bytes, and retained history memory is essentially unchanged.

Acceptance clarification: accepted edits keep per-edit undo semantics, even when the final document equals its initial value. The original net-zero criterion incorrectly assumed a frame-level transaction and is not implemented as such. Rejected/no-op paste preserves history (#374). Larger paste-payload scaling and broader retained-memory/latency acceptance are not established. Keep this investigation open rather than equating the scoped improvements with every proposed optimization.

[subinium]

v0.24.0 Disposition

Bulk splice, borrowed viewport rendering and removal of unwrapped idle document clones/maps are implemented. The measured 1 MB ASCII idle allocation count dropped from 24,549 to 54. Full-document undo snapshots remain: the cap-8 retained fixture is approximately 15.63 MB, essentially unchanged. A wrapped 1 MB CJK idle sample did not improve (6.654 -> 7.230 ms). Accepted edits retain per-edit undo semantics, including net-zero batches; rejected/no-op paste preserves history. The CI viewport property was corrected only after exact seed/minimal/no-undo reproduction, and two properties passed 20,000 generated cases. Keep the remaining memory/wrapped-work investigation open; do not infer a universal paste/latency speedup.

Implementation and evidence mapping: PR #433. Release v0.24.0 and its 20/20 release workflow are verified. This issue stays open only for the remaining acceptance/investigation described above; publication alone does not establish it.