perf(chat): batch live thinking rendering and bound timer updates

[RaresKeY]

Summary

Long reasoning streams repeatedly Markdown-parsed the entire cumulative thinking string and replaced its full innerHTML on every delta, while a requestAnimationFrame loop rewrote the elapsed header at display refresh rate. This patch coalesces live thinking to one plain-text DOM commit per 100 ms, performs one rich Markdown render when the block closes, synchronously flushes/cancels terminal paths, and bounds the timer to 250 ms normally or 1000 ms under reduced motion.

Target branch

• This PR targets dev, not main. All PRs land in dev; main is curated by the maintainer at each release.

Linked Issue

Fork staging: Issues are disabled on this fork. Rendered Issue body: A01-ISSUE.md.

Upstream publication packet uses Fixes #{{ISSUE_NUMBER}} after creating the focused Issue.

Part of odysseus-dev#4644.

Related to odysseus-dev#5588.

Publication sequence

This is slice 5/5 in the recommended performance publication order: A04 → A05 → A03 → A02 → A01. This branch is independently based on dev; it is not a cumulative stack branch.

Type of Change

• Bug fix (non-breaking — fixes a confirmed issue)
• New feature (non-breaking — adds new behaviour)
• Breaking change (changes or removes existing behaviour)
• Refactor / cleanup (behaviour unchanged)
• Documentation only
• CI / tooling / configuration

Checklist

• I searched open and closed Issues/PRs, including fix(ui): browser tab OOM and freeze during long agent interactions odysseus-dev/odysseus#4644, Thinking timer ignores reduced-motion; send button lacks an accessible name odysseus-dev/odysseus#5588, fix(ui): prevent browser OOM during long agent interactions odysseus-dev/odysseus#4661, fix(a11y): honour reduced-motion on the thinking timer + name the send button odysseus-dev/odysseus#5589, fix(chat): keep streaming TTS visible to errors odysseus-dev/odysseus#5857, and fix(chat): only leave agent mode when the model really has no tool support odysseus-dev/odysseus#5767; no open PR owns this rendering root cause.
• This PR targets dev.
• Changes are limited to static/js/chat.js and focused scheduler/lifecycle tests; no history virtualization, renderer replacement, document-editor work, or send-button accessibility change is mixed in.
• I actually ran the app and verified the long-stream behavior end-to-end in the disposable validation environment, including a manual huge-conversation check with no progressive lag.

How to Test

node --check static/js/chat.js
node --test tests/live_thinking_scheduler.test.mjs

python -m pytest -q \
  tests/test_live_thinking_scheduler_js.py \
  tests/test_chat_stream_scope.py \
  tests/test_bg_monitor_stream.py \
  tests/test_markdown_rendering_js.py \
  tests/test_streaming_segmenter_js.py \
  tests/test_strip_think.py \
  tests/test_redos_think_blocks.py \
  tests/test_llm_core_streaming.py

git diff --check e4fa4ae5dd1d709ce4168397bd1d200fec1b2494..HEAD

Prepared results:

• scheduler: 3 passed
• focused stream/thinking group: 37 passed, 1 warning
• JavaScript-area group: 202 passed, 4 skipped
• deterministic browser run: trailing sentinel exactly once, no page errors, bounded live-body mutation activity
• manual running-app huge-conversation check: no progressive lag
• final integrated sanity on agent/perf-all-integrated-a04-r2: scheduler 3/3, focused union 25/25, git diff --check clean
• fork staging CI previously passed CI, CodeQL, Secret scan, Workflow security, Container scan, and Trivy; Dependency Review was blocked only because Dependency Graph was disabled on the fork

Manual reviewer check:

1. Stream a 50k+ character reasoning block.
2. Confirm live output remains responsive and line breaks remain visible.
3. Confirm final Markdown renders correctly.
4. Confirm terminal/session transitions do not apply delayed rendering to a later session.
5. Optionally repeat under reduced motion.

Visual / UI changes — REQUIRED if you touched anything that renders

This changes behavioral rendering in the existing thinking block but adds no new styling or component.

• Screenshot or short clip — running-app behavior has been verified; attach a short long-stream clip or trace before marking the upstream PR ready for review.
• Style match — no CSS variables, colors, fonts, spacing, icons, or component markup are introduced.
• No new component patterns. The existing thinking block is reused.
• I am not an LLM agent submitting a bulk PR. Checked by the human publisher.

Screenshots / clips

Attach the final long-stream evidence before moving the upstream PR out of draft.

Risks and rollback

Open PR odysseus-dev#5857 also touches static/js/chat.js but owns an unrelated streaming-TTS error-scope fix. Rebase narrowly if that or other chat work lands first, then rerun the terminal-path regressions. Rollback is the two focused commits; there is no schema or persistence change.