perf(terminal): bound idle render costs for terminals and agent spinners

[onlynet]

Result / 用户结果

Idle terminals and waiting agent sessions no longer redraw the whole
window at display refresh rate. Previously a parked agent turn kept the
spinner (and therefore the window) at vsync rate, the cursor blink chain
ticked forever on unattended panes, every PTY read batch issued a frame,
and the 1 Hz watchdog ran a full-machine ps enumeration on the UI
thread (up to ~1 s of input latency per probe with large process tables,
measured in the Fleet fork on 2026-10-12). All of these costs now have a
local budget; a motionless terminal produces zero render frames.

Design / 设计边界

• Responsibility and affected modules: terminal/view.rs (cursor blink
  idle pause, output frame coalescing), workspace/agents.rs
  (SPINNER_TICK 80 ms chain, real-time phase), process_tree.rs
  (cached_activity_evidence, background thread + 250 ms shared
  snapshot), ai_agents.rs (identify/detect single-slot memo keyed by
  manifest generation), agent_detection/claude.toml (prompt_idle region
  widened to six bottom non-empty lines for the Claude CLI 2.x chrome).
• Each budget sits at its own source; there is no global frame
  throttling and no semantic change: blink restarts on any input, output
  storms neither keep the cursor lit nor add restart churn, spinner
  animation speed is unchanged (phase advances by real elapsed time).
• Threading: watchdog probes move off the UI thread; results are applied
  on the UI thread and stale batches are dropped by command identity.
• Compatibility: the memo never shortens the sampling cadence — the
  Fleet regression where caching short-circuited
  refresh_agent_screen_state and pinned agents at Running is recorded
  in the note and explicitly avoided (cache key includes the manifest
  generation so rule edits invalidate old conclusions).

Evidence / 验证依据

• cargo test -p nebula — full suite green on Windows; includes new
  unit tests for the pure decision functions
  (blink_silence_pauses_only_after_the_full_threshold,
  output-frame timing cases, spinner phase wrap).
• python scripts/check_architecture.py --base upstream/main — exit 0
  (587 counted source lines incl. the note).
• Hot-path changes: measurements cited from the Fleet fork sampling
  (2026-10-12): vsync-rate spinner redraws were the single largest
  idle-CPU item; watchdog stalls reproducible as periodic input latency.
  After: 12.5 Hz spinner chain, zero frames for motionless terminals,
  ≤125 fps output ceiling, UI thread never blocks on ps.
• Not run: macOS/Linux sides of the suite (Windows verification
  machine); no platform-specific code is touched beyond the shared
  watchdog path.
• Decision record:
  architecture/notes/nebula_app/gpui_shell/terminal/2026-10-12-idle-render-frame-budgets.md
  (a future render-snapshot cache stays a separate design gate and is
  deliberately not part of this PR).

Required Review / 必须确认

• I followed CONTRIBUTING.md, docs/architecture.md, and docs/project-constraints.md.
• I split responsibilities, not arbitrary line ranges; no duplicate behavior authority was added.
• python3 scripts/check_architecture.py --base <PR-base-commit> passes; budgets were not inflated to fit the change.
• Tests cover success and failure; platform/feature coverage limitations are stated.
• New messages use typed i18n IDs and matching placeholders; untranslated content has an explicit fallback.
• Governance changes include a counterexample, corrected contract, tests, and a maintainer-reviewed decision. (No governance changes in this PR.)

🤖 Generated with Claude Code