feat(web): render the reasoning stream as a collapsible thinking block

[Yevanchen]

Summary

• Render the reasoning stream (REASONING_MESSAGE_*) in the session chat instead of dropping it: a collapsed disclosure showing a shimmering "Thinking" label with a live character count while streaming, a static "Thought · N chars" header once prose/tools follow, expandable to the raw thought text.
• New reasoning segment kind in SessionLiveState (pkgs/ag-ui-session): REASONING_MESSAGE_START creates an assistant message keyed by the reasoning messageId, REASONING_MESSAGE_CONTENT appends to its reasoning segment; other REASONING_/THINKING_ events remain no-ops. Live-state schema extended to match.
• Web: convertSessionMessage maps reasoning segments to assistant-ui's native reasoning part; new AssistantReasoning component registered in MessagePrimitive.Parts (apps/web session-chat).

Why

Thinking models can reason for a long time before the first visible token — measured on the runtime catalog's DeepSeek default (deepseek-v4-pro): 148.7s to first visible character with 21.7k chars of hidden reasoning on a ~400-char writing task. The reducer previously no-opped every REASONING_* event, so that whole phase rendered as dead air; a user in YEF-891 replied "你吐字太慢了" mid-turn. The chat now shows continuous thinking progress instead of silence.

Verification

• Commands: bun test in pkgs/ag-ui-session (40 pass, includes 3 new reducer tests); tsc --noEmit clean for pkgs/ag-ui-session and apps/web; vp fmt / vp lint clean (pre-existing apps/api generated-tsconfig warning unrelated).
• Manual steps: rendered the real SessionThread + external-store runtime against fabricated live state in a dev harness (streaming and finished turns); verified shimmer + live char count while streaming, collapsed "Thought" header when done, expand/collapse of the thought text (screenshots on the linked issue).
• Not run: full platform e2e with a live DeepSeek session.

Impact

• User/API/contract changes: live-view only — a new reasoning segment kind in the SessionLiveState view model and its arktype schema. Persisted message history, GraphQL schema, and DB are untouched (reasoning is not persisted; it exists only during/within the live session view). Rollback consideration: live-state snapshots written with reasoning segments would fail schema parse on an older build.
• Generated files / GraphQL / DB / lockfile: N/A
• Env or config changes: N/A
• Risk and rollback: low; revert the commit. Empty reasoning parts are filtered by assistant-ui, so START-only messages fall back to the existing "Thinking…" empty indicator.

Review

• Closest review areas: pkgs/ag-ui-session/src/live-state-message-reasoning.reducer.ts (message keyed by reasoning id, ordering vs. tool/text messages); apps/web/.../session-message-parts.tsx (disclosure a11y/styling).
• Known trade-offs: reasoning is grouped as one block per turn (ACP emits a single thought stream id per turn), so interleaved think→tool→think phases accumulate into the same block rather than splitting; character count is shown rather than token count (token usage only arrives at turn end).

Context: YEF-891 — OpenCode + DeepSeek v4 preview feels silent/slow; companion driver fix langgenius/mosoo-agent-driver#81.