fix(ai): make annotation analysis reliable on mobile

[cha1latte]

What changed

• Prefetch the user's matching notes and highlights before model generation, so annotation requests cannot forget that the tool exists.
• Keep annotation-only turns scoped to getAnnotations, while preserving book retrieval and citations when the user explicitly asks for a comparison with the text.
• Give citation registration its own budget so citations no longer exhaust the ordinary tool-call allowance.
• Keep the Android/iOS display awake while any AI response is actively streaming.

Upstream dependency

• Depends on fix(ai): retrieve annotations by chapter and book position #719 for chapter-aware annotation filtering and pagination. Until fix(ai): retrieve annotations by chapter and book position #719 lands, current upstream still ignores chapterTitle, order, offset, and limit; book-wide prefetch works with the existing 20-item cap, but chapter scoping does not.
• Compatible with fix(ai): repair current chapter context #724. That PR improves current-chapter context freshness; this branch is based directly on current main and does not duplicate its changes.

Validation

• pnpm --filter @readany/core test — 80 files, 594 tests passed
• pnpm --filter @readany/app-expo test — 8 files, 18 tests passed
• pnpm exec tsc --noEmit -p packages/app-expo/tsconfig.json — passed
• Biome check on all changed files — passed (warning-level existing any diagnostics only)
• git diff upstream/main...HEAD --check — passed

The repository-wide pnpm lint remains red on the current upstream baseline because of pre-existing CRLF formatting and the vendored 1.9 MiB PDF worker; this PR does not modify those files.