## Problem
The late-wake detector added in #2559 suspends idle sleeping on perfectly healthy hosts. The visible symptom is this Warning firing periodically on stable machines:
> This host returned a 2ms idle wait at least 8ms late 2 time(s) in the last second; idle sleeping suspended for 5000ms
Demoting it to Debug would hide the symptom but not the cost: every one of those lines means the shard dropped idle sleeping for 5s and burned a full core for no reason. The detector is what was mis-tuned.
## Cause 1 — lateness was a count, not a rate
An idle loop performs **~400–500 sleeps per second** (2ms each, bounded by the 8ms wheel tick). The trip condition was `late > 1` across two consecutive one-second samples — a **0.4% tail-outlier rate**. A co-tenant burst, a page fault, or another process changing the system timer resolution clears that bar on a healthy host.
A host that genuinely cannot schedule the process — throttled burstable vCPU — returns *most* of its waits late. Signal and noise were two orders of magnitude apart, and the check sat in the noise.
Now gated on the proportion, with the absolute count kept as a floor:
```csharp
if (late <= _lateWakeThreshold) { _consecutiveBadSamples = 0; return; } // floor
if (late * 100 < sleeps * _lateWakePercent) { _consecutiveBadSamples = 0; return; } // rate
```
New `server.lateWakePercent` (default `10`). The floor is what keeps a window with only a handful of sleeps from tripping on a meaningless percentage; `server.lateWakeThreshold` keeps its existing meaning.
## Cause 2 — GC pauses were charged to the host
`dev-docs/debugging-event-loop.md` already documents that the GC collects preferentially **during idle sleeps** — that is the natural pause point it looks for. So the detector was systematically measuring the GC's chosen pause point and billing it to the host's scheduler. Not an occasional coincidence; a designed-in one.
```csharp
var collections = GC.CollectionCount(1);
NetState.WaitForCompletion(requested);
...
if (elapsed - requested >= Timer.TickRate && GC.CollectionCount(1) == collections)
```
Gen1 (which counts gen2 with it) rather than gen0 — gen0 pauses don't approach the 8ms `TickRate` bar anyway, and gating on them would discard useful samples. The second read short-circuits behind the overshoot test, so the common path costs **one** `GC.CollectionCount` per sleep: an internal counter read, single-digit nanoseconds, ~500/sec.
## Cause 3 — every backoff logged at Warning
Tiered to the escalation that already existed, since a single suspension is recoverable and not something an operator can act on:
| Backoff | Level |
|---|---|
| 1–2 | `Debug` |
| 3–5 | `Warning` (now includes the sleep count and "for the Nth time running") |
| ceiling | `Error`, unchanged |
| recovery | `Information` (new) |
Each backoff doubles the suspension, so every line is already a distinct escalation step — no further rate limiting needed.
## Drive-by
The `BackoffResetAfterCleanMs` reset only ran on the path to a *new* backoff, making it unreachable for a host that recovered for good — such a host never cleared its escalation or re-armed `_loggedBackoffCeiling`. It now runs on every health sample, which is also what makes the new recovery line reachable.
## Testing
Full solution builds clean, 0 warnings. No tests added: the state is private static in `Core` coupled to `_tickCount` with no injection point, and nothing covered it before — adding a seam purely to test it seemed worse than the gap. Happy to add one if reviewers disagree.
## Problem
`RunEventLoop` span through its body regardless of whether there was anything to do — ~10% of a desktop core for an empty shard, and ~70% of a core on a 3 vCPU VPS. A process that never idles is exactly what burstable vCPU plans throttle, which is how this surfaced: lag spikes that went away when the operator bought more cores. The spin also denied the GC its natural pause points, so memory climbed until a world save forced a collection — alarming in task manager, harmless in practice, and a recurring source of "is my server leaking?" reports.
## Result
Windows desktop, real world of **190,728 items / 33,158 mobiles**, no players, saves and prebake off, three consecutive runs:
| | Legacy spin | Idle sleeping |
|---|---|---|
| **CPU** | 10.42 – 10.50% of one core | **0.78 – 1.00%** |
| **Tick lag** (peak/15s) | 4–10 ms | 5–11 ms |
**~10× less CPU with tick lag unchanged** — the CPU came free rather than being traded for latency. Slower hosts gain proportionally more. Spin mode (`server.eventLoopIdleWaitMs=0`) independently gained **7× the iterations per core** (1.19M → 8.3M cycles/sec) from the ring's AcceptEx rework.
## How
The loop blocks in `NetState.WaitForCompletion` whenever every queue it drains is empty (all the drains are bounded, so leftovers keep it awake). Receive completions, new connections, and cross-thread `LoopContext.Post` (via the ring's sticky `Wake()`) are all in the wait set, so sleeping adds no latency to any of them. Only timer-driven logic sees wheel lag, bounded by the idle wait.
**Health is measured at the only place sleeping can cause harm.** A sleep is bounded by the time to the next wheel turn, so a correctly honoured sleep can never miss a deadline — the only failure mode is the host returning the wait late. That overshoot is measured on every sleep (one extra timestamp read; production's entire accounting cost), and an escalating backoff suspends sleeping when it persists. By construction, server work — saves, heavy staff commands, deep timer callbacks — cannot trip it, so the warning means exactly one thing: *the host is not scheduling the process promptly*, with two known remedies (dedicated CPU, or `=0`). Hosts with no high-resolution wait mechanism at all are detected once at startup and spin instead.
**CPS is removed.** `Core.CyclesPerSecond`/`AverageCPS` measured nothing actionable before and became actively misleading once the loop sleeps (the rate is set by the sleep, not by shard health). The admin gump's Performance page now shows the verdict instead: `Healthy` / `Sleep suspended (host)` / `Spinning (configured)`.
## Configuration
| Setting | Default | Meaning |
|---|---|---|
| `server.eventLoopIdleWaitMs` | `2` | Longest idle block. Measured across 1/2/4/8 ms, 2 is where the trade stops being free. `0` = never sleep: ~98% of a core, zero scheduling overhead — for large shards on dedicated CPU. |
| `server.lateWakeThreshold` | `1` | Idle waits the host may return a full tick late, per second, before sleeping backs off. Raise for jittery hosts; very high disables the backoff. |
## Diagnostics (compiled out by default)
`dotnet build -p:EventLoopProfiling=true` compiles in `EventLoopProfiler` — every hook is `[Conditional("EVENT_LOOP_PROFILING")]`, so normal builds contain zero profiling IL. The profiling build decomposes each second of wall time into **work (per loop phase) / sleep / GC pause / stolen residual**, keeps ~15 minutes of history in a ring buffer, and the `[LoopStats` command prints the last minute and dumps the full history to CSV. `dev-docs/debugging-event-loop.md` is the diagnosis guide (for humans and AI): what production already tells you, when to flip the profiling build, the signature table for host-steal vs deep-processing vs GC vs wake bugs, why dotnet-trace comes last, and the GC/RAM "leak" misconception.
## Verification
- 815 Server.Tests green; both build configurations compile.
- Docker echo harness green on epoll and io_uring (ping-pong mode); kqueue verified manually on an M1 Max.
- A/B measurements and per-change numbers: `measure/event-loop` branch.
## Notes
The full measurement harness and vendored ring sources used to develop this live on the [`measure/event-loop`](https://github.com/modernuo/ModernUO/tree/measure/event-loop) branch, kept for future loop work.
Kamron Batman