feat(annotations): non-modal, draggable, multi-thread reading companion

[ijbo]

What

Reworks the annotation "Ask AI" thread panel from a single anchored modal into a reading companion — so you can keep reading and ask in parallel, exactly as requested.

Before: the panel was modal (blocked the text behind it), one-at-a-time (opening a second closed the first), and couldn't be moved.

Changes

1. Non-modal + draggable

• Removed the full-screen backdrop from the Q&A path — the document stays readable and scrollable.
• Panels are draggable by their header (new grip handle), with cascade positioning and click-to-front.

2. Multiple threads in parallel

• Each annotation opens its own panel; opening another no longer closes the previous.
• Per-conversation state (streaming, web-search, attachments, model) moved from module globals onto each panel element, so threads stream concurrently without clobbering.

3. Right-side Threads dock

• An optional right rail holds docked threads as stacked cards; the document column reflows away from it (mirrors the existing AI-panel reflow) so nothing overlaps the text.
• Live thread count; collapses when empty; any thread toggles between floating ↔ docked via a header button. Full-width on mobile.

Verified live (browser preview)

• Two annotations → two parallel panels, 0 modal overlays ✅
• Cascade offset + z-index bring-to-front ✅
• Header drag moves a floating panel ✅
• Dock in → document reflows (app-container shrinks by the dock width); pop-out → restores ✅
• Close → cleanup + width restored ✅

Quality

• Vite build clean; smoke suite 22/22.
• Refactored the drag handlers to a single shared document listener (not per-panel) to avoid listener accumulation.
• ESLint: no new errors (pre-existing no-useless-assignment warnings untouched).

Honest limitation

Parallel threads share the globally-selected AI model — switching a panel's model switches the active model — because the local inference backend runs one model at a time. The UI supports independent threads; true per-thread concurrent models would need backend work.

🤖 Generated with Claude Code