Fix #6694 avoid LISTEN initialization in send-only Postgres channel paths

[michaelstaib]

Summary

• add regression coverage proving SendAsync should not initialize the listen connection in send-only scenarios
• split PostgresChannel initialization tracking into channel-listener and writer initialization
• update SendAsync to initialize only the writer connection, avoiding LISTEN startup in non-GraphQL background sender processes

Fixes #6694

Test Commands

• dotnet test src/HotChocolate/Core/test/Subscriptions.Postgres.Tests/HotChocolate.Subscriptions.Postgres.Tests.csproj --filter "FullyQualifiedName~SendAsync_Should_Not_Initialize_ListenConnection_When_SendOnly"
  • Repro before fix: failed on net8/net9/net10 with Assert.Single() Failure: The collection contained 2 items
  • After fix: passed on net8/net9/net10
• dotnet test src/HotChocolate/Core/test/Subscriptions.Postgres.Tests/HotChocolate.Subscriptions.Postgres.Tests.csproj --no-build
  • Passed on net8/net9/net10 (Failed: 0, Passed: 74, Skipped: 0)

[github-actions]

Fusion Gateway Performance Results

Simple Composite Query

	Req/s	Err%
Constant 2 (50 VUs)	3288.48	0.00%
Ramping 2 (0-500-0 VUs)	3886.61	0.00%

Response Times

	Min	Med	Avg	P90	P95	Max
Constant 2	1.13ms	12.25ms	14.93ms	24.29ms	34.00ms	257.57ms
Ramping 2	1.36ms	51.35ms	56.44ms	98.59ms	139.93ms	283.75ms

---

Deep Recursion Query

	Req/s	Err%
Constant 2 (50 VUs)	784.90	0.00%
Ramping 2 (0-500-0 VUs)	913.22	0.00%

Response Times

	Min	Med	Avg	P90	P95	Max
Constant 2	12.09ms	54.85ms	61.54ms	78.96ms	97.47ms	736.57ms
Ramping 2	2.92ms	202.71ms	225.79ms	467.05ms	521.41ms	798.99ms

---

Variable Batching Throughput

	Req/s	Err%
Constant 2 (50 VUs)	5954.01	0.00%
Ramping 2 (0-500-0 VUs)	5129.95	0.00%

Response Times

	Min	Med	Avg	P90	P95	Max
Constant 2	0.15ms	7.84ms	8.30ms	13.76ms	16.86ms	59.93ms
Ramping 2	0.19ms	40.35ms	44.77ms	83.07ms	107.62ms	229.04ms

---

Runner 2 = benchmarking-2

Run 22490365189 • Commit 0f2f83b • Fri, 27 Feb 2026 14:43:27 GMT