fix(web): keep user messages stable

[Yevanchen]

Summary

• Keep optimistic and canonical user messages on one stable assistant-ui identity.
• Bypass output pacing for content belonging to a role: user message.
• Add regressions for stable user IDs and immediate delivery of long user input.

Why

The canonical WebSocket echo used a different message ID from the optimistic message, so the user bubble was deleted and remounted. Separately, TEXT_MESSAGE_CONTENT has no role of its own and entered the assistant-output scheduler, making the remounted user input appear to stream.

Verification

• Commands:
  • just test-file apps/web/tests/session-message-convert.test.ts (8/8)
  • just test-file apps/web/tests/session-stream-render-scheduler.test.ts (22/22)
  • just tc-package @mosoo/web
  • just test-package @mosoo/web (217/217)
  • bash ./init.sh acceptance
  • just react-doctor-diff
  • git diff --check
  • just check (1081 pass, 42 skip, 1 unrelated failure)
• Manual steps: sent a 308-character Chinese prompt through the local UI and sampled before/after the server echo; exactly one full user bubble remained mounted as user:0 at every sample.
• Not run: no production deployment. The full gate's sole failure is the unchanged ACP filesystem test expecting /var/... while macOS returns /private/var/...; the focused test reproduces the same path-normalization mismatch.

Impact

• User/API/contract changes: user input stays fixed; only assistant output is paced. No API or wire-contract change.
• Generated files / GraphQL / DB / lockfile: N/A
• Env or config changes: N/A
• Risk and rollback: limited to Web transcript identity and render scheduling; revert 4981fbc580 to roll back.

Review

• Closest review areas: assistant-ui external-store message identity; session stream scheduler role tracking.
• Known trade-offs: user IDs are transcript-position based inside the rendered session; source IDs remain unchanged outside the UI adapter.