Flaky timing-sensitive sync-service tests: ShapeCacheTest await/startup timeouts under CI load

[robacourt]

Summary

A small family of timing-sensitive tests fails intermittently in the sync-service CI workflow — hopping between Postgres matrix versions, passing on rerun with no code change, and burying real signal for PR authors. Two recent PRs (#4711 and its review) each spent effort separating these from genuine failures.

Evidence

Electric.ShapeCacheTest — "await_snapshot_start/4 should wait for consumer to come up"

• Failed on pg15: https://github.com/electric-sql/electric/actions/runs/29346393580/job/87131112285 — Task.await(..., 5300) timeout
• Passed on rerun of the same commit (all matrices green)
• Failed on pg17 two commits later: https://github.com/electric-sql/electric/actions/runs/29348549229/job/87138530389 — same test, same timeout
• The failing branch's changes don't execute in this test (verified: the changed code is HTTP-serve-path only, and the branch's heavy async: false integration tests run after the async phase this test runs in — no load coupling)

Electric.ShapeCacheTest — "get_or_create_shape_handle/2 against real db crashes when initial snapshot query fails to return data quickly enough"

• Failed on main: https://github.com/electric-sql/electric/actions/runs/28917768297

Electric.Plug.TraceContextPlugTest — "electric member is found among other tracestate members"

• Failed on main: https://github.com/electric-sql/electric/actions/runs/28605359873

ShapeLogCollectorTest — FlushTracker timing out in Repatch.cleanup

• Reported during review of fix(sync-service): terminate serves whose clients stop accepting data #4711 (first CI run of that branch)

Common signature

• ~5s Task.await timeouts on tests that race consumer/snapshot startup
• Postgres connection churn in the surrounding logs (tcp connect timeout, econnrefused, tcp recv (idle): closed) suggesting the shared PG instance is briefly saturated by concurrent async tests
• Not reproducible locally on fast machines; version-hopping across the PG matrix

Suggested directions

• Widen or parametrize the tight awaits in the affected tests (they're asserting eventual startup, not 5-second startup)
• Or address the underlying PG saturation (connection pool sizing / readiness in the test env)
• At minimum, tag the known-flaky set so authors and reviewers can distinguish them from real regressions without re-deriving this analysis

---

Filed while triaging CI on #4711, where these flakes twice muddied the signal.