Reduce orchestration worktree freeze pressure

[Juliusolsson05]

Summary

• Add main-side backpressure for renderer orchestration requests so bulk list/read/mark calls cannot flood the renderer bridge.
• Make orchestration list/status cheap and make read-agent/read-run-outputs tail-bounded instead of materializing full transcripts before slicing.
• Reduce git/worktree fanout with a process-wide git subprocess limiter, longer/aliased worktree caches, fewer status subprocesses, and cheaper submodule dirty checks.
• Stabilize WorktreesBar polling so runtime churn does not restart worktree scans and overlapping refreshes are coalesced.

Root cause

The failure mode is renderer event-loop starvation. Orchestration MCP coordination routes through the renderer-owned workspace model; under many agents, wait/list/read calls were scanning every child transcript while worktree UI/session paths also spawned many git subprocesses and pushed broad runtime updates. Main then reported the bridge as unhealthy because it was waiting on a renderer that was busy doing synchronous coordination/UI work.

Validation

• git diff --check passes.
• npm exec tsc -- --noEmit --pretty false -p tsconfig.node.json is blocked by existing missing local package declarations and unrelated existing type errors.
• npm exec tsc -- --noEmit --pretty false -p tsconfig.web.json is blocked by generated node declaration outputs missing after node typecheck fails.
• npm exec eslint attempted to fetch eslint from npm and failed because network is restricted/no local eslint binary is installed.