perf(daemon): make scheduling passes linear in the number of tasks

[arcanis]

Issue: a trivial task across 1000 workspaces took 16-20s. After every command, including each output line, the coordinator rescanned every task and allocated keys for each prerequisite check.

Fix:

• Passes iterate prepared tasks once, with precomputed satisfied/failed sets and borrowed-key lookups.
• The failure scan is skipped when nothing failed.
• Commands that can't change task states don't trigger a pass.
• Standalone daemons don't register taskfile watchers.

1000 tasks now take ~1-2s. Stacked on #376.

---

Note

Medium Risk
Scheduling and prerequisite propagation logic changed on the critical coordinator path; incorrect skips or set membership could delay spawns or miss failure propagation, though behavior should be equivalent if the command classification is complete.

Overview
Improves daemon performance for large monorepos by cutting redundant scheduling and expensive startup work.

The coordinator skips drain_ready_tasks after read-only commands (notably TaskOutput and other subscription/query/file-watch paths) so high-volume output no longer triggers full graph scans.

find_ready_tasks / find_tasks_to_fail are rewritten to build satisfied/failed sets once per pass and check prerequisites with borrowed keys instead of allocating a ContextualTaskId per edge; failure propagation returns immediately when nothing failed.

Inline (standalone) daemons detect port_tx and disable taskfile hot-reload watching (TaskfileWatcher::set_enabled(false)), avoiding O(workspaces) FSEvents registration that was dominating startup.

^{Reviewed by Cursor Bugbot for commit e5b5e7f. Bugbot is set up for automated code reviews on this repo. Configure here.}