Real app reproduction on base 70e5e384da, Linux x86_64, Node v24.21.0
Isolated checkout /home/ubuntu/wt/issue-18799-before; vp run dev
[dev-runner] serverPort=16664 webPort=8624 baseDir=/home/ubuntu/wt/issue-18799-before/.t3
Seed: copied stopped sandbox statev2.sqlite, synthetic local-command ACP instance only; Codex/Claude disabled in sandbox.
Browser: T3 collaborative preview, 1280x800, no emulation.
Fixture: original packages/effect-acp/test/fixtures/acp-mock-peer.ts, prompt modified to emit 40 x/large extension notifications with distinct 8 MiB strings then idle state update; no raw stream reader.
Action: New thread > select project > Message='Emit synthetic retention notifications' > Submit.
Attempt 1: fixture used Effect.log on stdout, corrupting ACP JSON. App reported 'Provider turn failed'; server AcpProtocolParseError: Expected ',' or ']' after array element in JSON at position 3. This was a fixture error, not the reported bug.
Attempt 2: moved fixture diagnostic to stderr; prompt emitted 40 notifications, then original fixture waited on permission/elicitation. App displayed Input. Server remained connected. GC via local Node inspector: heapUsed=471616840, external=60080871.
Attempt 3: removed unrelated permission/elicitation calls from fixture; new thread emitted another 40 x 8 MiB and idle update. UI displayed Working/Thinking with no visible messages (unknown extension notifications have no rendered message). Server remained connected; post-GC heapUsed=725143024, external=60081417. Increment over attempt2=253526184 bytes (~241.8 MiB). This demonstrates retained large payloads on the actual app server path. It is synthetic traffic, not a real paid provider session.
The direct protocol measurements in before/after JSON files isolate the queue effect; total server heap contains other app state and should not be treated as exact notification sizes.