Performance

Budgets come from REQUIREMENTS (NFR-1, NFR-3). make perf (scripts/perf.sh) builds a synthetic fleet of local repositories, runs the real binary against it and fails if a budget is missed. Everything happens in a temp dir with isolated HOME, XDG and git config.

Result (500 repositories, Linux, x86-64, 4 workers)

Metric	Measured	Budget
Startup until the UI is listening	68 ms	< 300 ms
First full cycle (all 500 cloned repos synced)	152 s	none (dominated by the 250 ms first-sync stagger: 500 × 0.25 s = 125 s)
Peak RSS during the cycle	23 MB	< 50 MB
Idle RSS	23 MB	< 50 MB
Idle CPU (30 s window)	0.000 % of one core	< 0.5 %
Dashboard response with 500 repos	16 ms	< 500 ms
Dashboard fragment refresh	21 ms	none
Dashboard HTML size	304 KB	none
`state.db`	264 KB	none

"Idle" is the window right after the first cycle, when every repository's next sync is about 30 minutes away. During continuous operation the cost is the git subprocesses of whichever repositories are due (one ls-remote, usually no fetch because the remote digest is unchanged).

What the benchmark found and changed

First-sync stagger of 2 s per repository meant a 500-repo fleet needed ~17 minutes before the last repository's first sync. Reduced to 250 ms; real hosts are additionally paced by the rate limiter, so this only avoids a burst of process spawns on the local machine.
Local remotes were being rate limited like a remote host (1 request/s). They now bypass the host limiter (there is no server to protect); the worker pool still bounds concurrency.

Known characteristics and limits

The dashboard renders every repository on one page (304 KB for 500). It stays fast, but at several thousand repositories pagination will be needed (not implemented).
Hosted platforms are the real bottleneck: with the default pacing (1 request/s per host, burst 5) a first sync of 500 GitHub repositories takes on the order of ten minutes by design.
Numbers were taken on one machine; absolute values will differ, the budgets leave an order of magnitude of headroom.
Re-run make perf after changes to the scheduler, store, UI templates or git runner.