[tobias-tengler]

For every request, the Fusion timeout middleware creates a timeout CancellationTokenSource with its own timer, a linked source over request abort, timeout and schema shutdown, and a registration on the schema shutdown token, which all requests share.

This PR rents a pooled RequestTimeoutSource that keeps one token source and one timer and registers on the schema shutdown token once, so a request only registers on its abort token and arms and disarms the timer. The timeout keeps its own flag, so a request abort or schema shutdown is still not reported as a timeout, and a timer callback dispatched for an earlier request cannot cancel a later one. Streams keep the source until their cleanup as before.

Method	Version	Allocated (round 1)	Allocated (round 2)	Mean	Delta
POST through the gateway	before	22,348 B	22,344 B	22,346 B
POST through the gateway	PR	21,781 B	21,798 B	21,790 B	−556 B

Allocations per request from an allocation probe: a TestServer gateway answers a { value } POST through an in-process source schema handler that yields once, 20,000 sequential requests after 5,000 warmup requests, GC.GetTotalAllocatedBytes(precise: true) across all threads (client, server, gateway and source handler), two rounds. The probe ran with the series applied cumulatively on main, so the before row is main with #10618, #10619, #10620 and #10621.

For context, the k6 constant screen of this change together with the other request pipeline changes (#10618, #10619, #10620, #10621 and #10623) and one further change that is not proposed, A B B A per gateway (3 × 30 s runs per invocation). RPS is the median of the six runs, CPU per request the mean of the six runs, p95 and p99 are per invocation.

Gateway	Version	RPS	CPU per request	p95	p99
Composite schema (net10.0)	main	4,461	590.6 μs	19.47 / 19.36 ms	24.87 / 25.15 ms
Composite schema (net10.0)	Bundle	4,434	586.1 μs	19.21 / 19.45 ms	24.96 / 24.98 ms
Apollo Federation (net11.0)	main	4,109	566.6 μs	19.96 / 19.96 ms	25.50 / 25.48 ms
Apollo Federation (net11.0)	Bundle	4,104	563.3 μs	19.84 / 19.91 ms	25.59 / 25.05 ms

The change saves about 556 B per request, far beyond the probe's run-to-run noise of about ±50 B, and requests no longer register on the shared schema shutdown token. The bundle shows no regression: RPS (−27 and −5), CPU per request (−0.76% and −0.58%) and latency all stay within the A/A noise measured on main (σ about 65 RPS for the composite schema gateway and 40 RPS for the Apollo Federation gateway, CPU per request CV 0.5 to 0.8%), so the end-to-end gain is not measurable on its own.

[github-actions]

Patch coverage

97.3% of changed lines covered (71/73)

File	Covered	Changed	Patch %
…/Fusion.Execution/Execution/Pipeline/RequestTimeoutSource.cs	59	61	96.7% 🟡
…/src/Fusion.Execution/Execution/Pipeline/TimeoutMiddleware.cs	12	12	100.0% 🟢

Uncovered changed lines (JSON)

{
  "sha": "933d7ef799b4693f2adec4025a8700c0498f48a1",
  "files": [
    { "path": "src/HotChocolate/Fusion/src/Fusion.Execution/Execution/Pipeline/RequestTimeoutSource.cs", "ranges": [[121, 122]] }
  ]
}

---

Project coverage: 59.0% (311605/527913 lines)