fix(direct): sentence-window context + remove per-keystroke re-reads (Outlook canvas reset)

[willwade]

Closes the Outlook canvas-reset problem (Heide's video report).

Root cause

The per-keystroke UIA selection watch (#56, RFC 0019 clause 6) re-read the target on every selection change. Outlook fires TextSelectionChanged on every injected character and its UIA provider returns inconsistent full-document reads (signatures, formatting, timing) — the shadow-compare against the full text could never reliably match, so every keystroke triggered a seed_buffer → full model rebuild → visible canvas reset. BlueMail's "stutters but usable" was the same mechanism at lower severity.

v5 never had this problem because it never re-read the target within a field — its internal EDIT control accumulated whatever was typed through Dasher, full stop.

Fix (Heide's suggestion, v5-aligned)

1. Sentence-window seeding: target reads trimmed to the sentence around the caret (last sentence boundary, max 200 chars back) before seeding. The LM only needs local context; a small seed makes the shadow-compare stable — during typing, the sentence and the engine buffer grow together character by character, so echoes match and are skipped.

2. Selection-change watch removed entirely. Context seeds once on focus/app switch (the focus hooks), then the engine buffer accumulates internally (v5 parity). No per-keystroke re-reads. Clicking into a different field/app fires the focus hooks and re-seeds with that field's sentence.

3. The long-document seed cap is superseded by the sentence window for the target path (always ≤ 200 chars); EditorSeedPolicy remains for the editor pane.

Tests

10 new SentenceWindowTests: boundary detection (. ! ? \n ; :), the typing-flow sync invariant (sentence grows in lockstep with engine buffer → shadow-compare skips), multi-sentence clicking, caret-at-boundary → empty context, max-lookback cap, empty inputs. Full suite 108/108.

Manual: publish-test built — test in Notepad and Outlook (type a word, verify no canvas reset per keystroke; switch between fields; the keyboard_debug.log shows sentence "..." (N of M chars) lines).

Confidence Score: 5/5

The PR appears safe to merge, with no outstanding previous findings or actionable new failures identified.

Summary

• Adds bounded, surrogate-safe sentence-window extraction and focused unit coverage.
• Preserves same-field caret and external-edit synchronization through the UI Automation selection watch while avoiding rebuilds for matching injection echoes.
• Adds a manual re-anchor action and ensures selection-watch lifecycle and failed-arm retries are handled correctly.
• Removes target sentence content from diagnostics and safely reseeds whitespace-only divergences.
• Advances DasherCore for aggregated invalid-UTF-8 training-data warnings.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    Target[Focused target field] --> Read[Read text and caret]
    Watch[UIA selection-change watch] --> Read
    Manual[Manual re-anchor] --> Read
    Read --> Window[Extract bounded sentence window]
    Engine[Native engine output] --> EngineWindow[Extract matching sentence window]
    Window --> Compare{Sentence comparison}
    EngineWindow --> Compare
    Compare -->|Exact match| Offset[Cheap offset update or skip]
    Compare -->|Whitespace-only difference| Reseed[Reseed current sentence]
    Compare -->|Content difference| Reseed
    Offset --> Canvas[Preserve prediction canvas]
    Reseed --> Canvas

Loading

_{Reviews (5) · Last reviewed commit: "chore: pin DasherCore to v0.2.24 + UTF-8..."}