perf(1346): round 2 -- the statusbar and clipboard costs the first budget could not see

[accesswatch]

Re-examining #1346 found a second layer larger than what round 1 fixed. The round-1 budget test stubbed _refresh_title — and _refresh_title was smuggling a full statusbar refresh onto the synchronous typing path.

What each keystroke still did after round 1

• The statusbar refreshed twice — once dragged in synchronously by the title refresh, once from the key-up handler — and its cells each re-marshalled the buffer: word count (full copy + full text pass), line/column (full copy + scan to caret), progress (full copy), page (full copy + scan). ~9 full-document copies and 2 full scans per character.
• The abbreviation expander opened the Windows clipboard on every keystroke. The clipboard is a shared, single-owner cross-process lock; screen readers and clipboard managers poll it, and clipboard_retry deliberately retries up to 10 × 20 ms on contention — per character typed. Of everything in either round, this is the most direct candidate for the reported NVDA/JAWS lag.
• SetTitle and the tab label were re-set to identical strings per keystroke, each firing an MSAA/UIA name-change event.

The fixes

• try_expand gains clipboard_provider: the clipboard is read only once an abbreviation has matched and its expansion references ${clipboard} — approximately never. Lazy memoized fetch threaded through the user library, the Quillin-contributed library, and the fill-in fields path.
• _refresh_title splits into the cached _refresh_title_bar (native calls skipped when unchanged) plus the statusbar half. The typing path uses the title-bar half; the statusbar catches up in the deferred pass. Caret activity coalesces its refresh (90 ms restarting timer) — holding an arrow key costs one refresh when the caret stops, not one per repeat.
• Display-only reads (statusbar cells, table-transition check) use the new _document_text_for_display() — document.text is the same string GetValue would marshal, already in memory, zero copies. The bug: quill crashes when running from source, post startup wizard, and when pressing ctrl+f4. #269 dead-widget contract is kept via an insertion-point liveness probe.
• Document.set_text compares length before content.

Measured (same modelling basis as round 1)

Document	After round 1	After round 2
20 KB	2.1 ms/keystroke	0.014 ms	156x
200 KB	22.4 ms/keystroke	0.26 ms	86x
1 MB	117.9 ms/keystroke	1.28 ms	92x

— before counting the removed clipboard contention, which no model captures.

Gates

• Budget test: 10 → 16 assertions (clipboard laziness at three levels, title/statusbar split, caret coalescing, zero-copy display reads, title caching).
• Full tests/unit + tests/stability: 13089 passed, 31 skipped, 0 failed.
• GATE-11: four rebaselines with dated rationale (_rebaseline_2026_08_13_1346_round2). ruff + scoped mypy clean. Docs (CHANGELOG, PRD §11, user guide) updated, artifacts re-rendered, parity gate passes.
• Navigation-test fixtures now keep the stub document in sync with the stub editor, as the real app does on every EVT_TEXT — the old fixtures modelled a desync that cannot occur.

🤖 Generated with Claude Code