Follow-ups from #702: a raised-memlock CI step for the celeris#592 rig, with the io_uring stall assertion

[FumingPower3925]

Follow-ups from PR #702 (Refs #684), deferred under the maintainer's two-round review cap (2026-09-27). #702 changes two test files only; this item is CI wiring it deliberately left out (its body lists it under "Not in this PR", and the #684 census records it: #684 (comment)).

1. No CI step runs the celeris#592 rig with memlock raised and CELERIS_REQUIRE_IOURING_WORKERS=1 (CodeRabbit, minor, test: run the epoll sendfile e2e tests, and let CI forbid the celeris#592 rig's io_uring skips (celeris#684) #702 (comment); verified at dc14ef5). test: run the epoll sendfile e2e tests, and let CI forbid the celeris#592 rig's io_uring skips (celeris#684) #702 routes both io_uring skips of TestAdaptiveSettledRouteRetime592 through skipOrFailIOUring592, which fails when the variable is set, but no step sets it for this test. The thread's premise is off: the test is in package celeris (the repository root, adaptive_settled_retime_linux_test.go), so the adaptive job (./adaptive/...) never runs it, and adding the variable to that job's env, as suggested, would change nothing for it. It runs only in the Unit job's root step, at the runner's 8 MiB, without -v, where its three io_uring subtests skip silently (ci.yml has 0 matches for Retime592; positive control TestFlapConnsPerRing: 3). The fix is a step (or an entry in an existing raised-memlock step) that runs -run '^TestAdaptiveSettledRouteRetime592$' . with sudo prlimit --memlock=unlimited, -v, CELERIS_REQUIRE_IOURING_WORKERS=1 and a PASS/SKIP tally (6 subtests, 0 SKIP lines). Per the CI: with io_uring unavailable, 16 ./adaptive tests skip and the adaptive job passes even under CELERIS_REQUIRE_UPSWITCH=1 #684 census it is worth adding together with item 2, since on its own it repeats the adaptive dispatch never re-times a settled route: a store-backed handler that turns slow runs inline on the engine worker forever (#493 item 4, measured) #592 check the epoll half already makes.

2. The rig's io_uring half does not catch a celeris#593 revert (from the CI: with io_uring unavailable, 16 ./adaptive tests skip and the adaptive job passes even under CELERIS_REQUIRE_UPSWITCH=1 #684 census, recorded here so the CI step and the assertion land together). With the snapshotH1Deadlines TryLock reverted to a blocking Lock, the io_uring half still PASSES 3/3: it prints a stalled fraction of 0.327-0.340 against 0.000, but asserts only a median ratio. Making it assert the stall it prints is the owner's call on the bar, given celeris#622's 0.145 queued-fraction outlier.

Also on that thread: ast-grep flags t.Fatal(msg + " -- CELERIS_REQUIRE_IOURING_WORKERS=1 forbids skipping") as log injection (CWE-117). It is a false positive: msg is the test's own format string and a local error, not request data, and the sink is testing.T.