fix(chat): cross-tab isolation, stream/history consistency, and reasoning replay

[Luhaozhu]

Summary

This PR lands a batch of stability and correctness fixes for the chat experience, discovered during full end-to-end functional testing, plus first-class support for models that deliver reasoning through the structured reasoning_content channel.

Multi-tab session integrity

• The chat tree in localStorage is now written with per-chat granularity merge (newest updatedAt wins) instead of whole-tree overwrite, and tabs sync via the storage event. Two tabs chatting in parallel no longer clobber each other's conversations.
• The current-chat pointer is tab-private (sessionStorage first, localStorage as fallback for fresh tabs), so refreshing one tab can no longer hijack another tab's restore target.
• Following a background run's SSE stream is guarded by Web Locks keyed on the run id — only one tab replays a given run, eliminating duplicated / half-rendered bubbles and cross-conversation bleed.
• Restoring a session keeps the previous chat id even when the backend does not know it yet (e.g. the first message is still streaming), so duplicating a tab no longer mints a phantom conversation.

Stream vs. history consistency

• Each persisted tool call now records a content_offset (the length of the accumulated answer when the call was emitted). History rebuilds interleave text and tool cards in the original streaming order — a refreshed page now shows exactly what the live stream showed.
• Visible answer text preceding a <think> open tag is no longer dropped when rebuilding history segments.

Structured reasoning channel

• Thinking deltas are persisted as interleaved <think>…</think> blocks (same storage format as inline-thinking models), so the thought process replays after a refresh.
• When the first model round finishes without any </think> in its text, the backend emits a structured-reasoning protocol marker; the frontend then reclassifies provisionally-buffered text back into the answer body. This fixes answers being trapped inside the thinking block when a reasoning-capable model produces no reasoning for a trivial prompt.

Other fixes

• Edit-and-resend during streaming stops the in-flight run and resends, instead of silently ignoring the click; opening/cancelling the editor no longer force-scrolls to the bottom.
• Wheel-up during streaming immediately detaches follow-to-bottom (no more fighting the auto-scroll).
• Per-answer duration is persisted (extra_data.duration_ms) and survives refresh; display unit unified.
• Share-page printing expands the inner scroll container so the full conversation prints (previously only page 1).
• Renaming with empty input keeps the original title; manual renames set title_manually_set and are never overwritten by auto-summary; renames made mid-stream sync to the backend once the session exists.
• Conversation search and favorite-card previews operate on the visible answer only — thinking content (including raw </think> fragments) is excluded from matches and snippets.
• Attachments uploaded while a stream is running survive stream completion.
• Message loading retries with backoff after transient failures instead of sticking on the skeleton.
• Knowledge-base document search is now a server-side whole-library search (title/filename ilike) with pagination over the filtered set.

Testing

• Automated browser (Playwright) end-to-end runs against a local deployment: dual-tab concurrent chats, duplicated-tab refresh, tool-calling conversations compared live vs. refreshed, print-to-PDF page count, injected 503s for the retry path, reasoning-mode prompts with and without thinking output.
• Backend unit suites covering chat repository search, artifacts, and KB pass; frontend type-check and stream-segment regression script pass.