fix(app): keep slow git reads from filling the request queue

[Hona]

The request queue from #47441 is a plain FIFO with 4 slots. A Desktop netlog shows what sits in those slots during a session-tab mount:

Endpoint	Server time
GET /api/vcs	p50 1789 ms, max 2468 ms
GET /api/worktree	p50 461 ms, max 2243 ms
GET /api/session/:id, /message, /permission, /form, catalogs	p50 2–25 ms

A mount fires ~21 requests; two of them run git and can hold slots for over two seconds each. With several tabs mounting or a reconnect re-syncing directories, git reads take all 4 slots and the reads that actually paint the session sit behind them — which is exactly what server thrashing detected was logging (inflight: [GET /api/worktree…, GET /api/worktree…], queued: [GET /api/health]).

flowchart LR
  subgraph before [Before: FIFO x4]
    B[vcs · vcs · worktree · worktree] --> Q[queued: session, messages, permissions…]
  end
  subgraph after [After: slow share = 2 of 4]
    A1[vcs · vcs] --- A2[session · messages · …]
    A3[worktree · worktree] -.waits for a slow slot.-> A1
  end

Loading

Change — packages/app/src/runtime/server/request-queue.ts

• slowRequestPaths = ["/api/vcs", "/api/worktree"] (prefix match, so /api/vcs/diff, /api/vcs/branches included); requestQueueSlowLimit = 2.
• Slow requests may hold at most 2 of the 4 slots. Order stays FIFO otherwise: a slow request only yields its turn while the slow share is full, so nothing starves and fast reads never change order among themselves.
• The wait list is scanned for the first eligible entry on each release instead of popping the head.

const canStart = (entry: Entry) => {
  if (inflight.size >= limit) return false
  if (!entry.slow) return true
  return [...inflight].filter((item) => item.slow).length < slowLimit
}
const release = (entry: Entry) => {
  inflight.delete(entry)
  const index = waiting.findIndex((item) => canStart(item.entry))
  if (index === -1) return
  waiting.splice(index, 1)[0]?.start()
}

Total concurrency is unchanged (4 + the event stream), so the Chromium per-origin limit reasoning in #47441 still holds.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.