Let the keyboard alone decide a sentence start

[enaboapps]

Closes #910. Part of the casing plan in #914.

What changed

• The private worker query carries sentence_start, the keyboard's pending automatic capital: a full stop, exclamation or question mark it typed, not yet followed by a letter, and not cancelled by pressing Shift. In-process pipe message only; no persisted schema change.
• The worker capitalises a suggestion's first letter from that flag alone. Its own guess, from punctuation seen in the buffer or the bare start of the buffer after a reset, is removed along with the sentence_start field of the extracted context. Prefix and preceding-text extraction are unchanged.
• The cached-batch key includes the flag, so a change in the keyboard's state refreshes the row.
• The old test that capitalised from "Done! " in the buffer now asserts that buffer punctuation alone gives "water" and the keyboard's flag gives "Water". Modifier combinations at a sentence start are tested with the flag set.
• docs/word-prediction.md explains that the keyboard alone decides.

Behaviour change worth knowing: at the true start of a document, with nothing typed and no punctuation, the keyboard has no signal, so the first word is not auto-capitalised. Shift once gives "Water". Before, the worker guessed a capital at the bare start of its buffer, which was right at a document start and wrong mid-line.

Validation

Node 24.13.0, Rust 1.97.1:

• npm run lint, npm test (244 Vitest tests and 5 Node tests) and npm run build passed. No frontend files changed.
• cargo fmt --check and cargo clippy --all-targets -D warnings passed.
• cargo test passed: 567 tests.

🤖 Generated with Claude Code