Restore the BUILD_SAMPLES option; state BUILD_MOQTEST; fix the XLOG smoke test

[gmarzot]

Restores the BUILD_SAMPLES option that sync #375 dropped, states BUILD_MOQTEST explicitly at every configure site, and fixes the XLOG category smoke test.

Reworked from r1 per review: the BUILD_SAMPLES removal is reverted and left for after the moqtest rewrite lands. BUILD_TESTS stays too — see the last section.

The sync fix

Upstream D116974940 inserted option(BUILD_MOQTEST ...) into its own standalone/CMakeLists.txt. Upstream's copy has never declared BUILD_SAMPLES, so that is correct there; ours declares it on that exact line, and the merge in 065e3784 replaced our line instead of keeping both.

Standalone saves BUILD_SAMPLES, forces it off while the fetched dependencies configure, then restores it. With no declaration the saved value was empty, the restore set "", samples/ stopped building, and the tarball for 91f2387 lost moqdateserver and moqtextclient. openmoq/moqx#620 failed relay_chain and relay_hops_cycle with not found or not executable: .../bin/moqdateserver.

BUILD_MOQTEST=ON is now passed at all four configure sites rather than inherited from the option default, so a sync dropping a declaration cannot silently change what a build produces.

The smoke test runs on bash 3.2

xlog_category_roots has reported Passed 0.01 sec on macOS since it landed in #312, without running anything. The marker map used declare -A and assert_scoping used a local -n nameref; both need bash 4, macOS ships 3.2, and without set -e the failures were silent — assert_scoping never ran, fail stayed 0, exit 0.

Parallel arrays and a marker_for lookup replace them. Verified under bash 3.2.57 on a macOS box with stub binaries, exercising the arrays, the nameref-free assert_scoping, and the lsof readiness path where ss is absent.

This means macOS will genuinely run this test for the first time. Worth watching that job.

The smoke test is faster

Every marker the test looks for is emitted during connection setup, and the client runs until killed, so the old shape spent essentially all its time waiting. Four sessions of sleep 1 plus timeout 3 account for 17.2s of a 17.44s CI run.

A readiness poll on the server's UDP socket replaces the sleep. The client is backgrounded and killed rather than run under timeout(1), which macOS lacks.

	time
before	18.42s
after	4.0 - 4.9s across three runs

Same ASAN build, same three PASS and three SKIP lines. In CI terms 17.4s to roughly 4s.

Worth being clear this is not a CI speedup: the linux job is 21m28s, of which 18m48s is the build and 72s is all 1555 tests. What it fixes is an outlier — this one test took 24% of test time and ran 11x longer than the next slowest.

On dropping BUILD_TESTS

Agreed in principle, but it breaks moqx today and the fix belongs in moqx.

moqtest_utils is indeed always built now. It is not the only thing moqx takes from the tarball though — the googletest acquisition is inside if(BUILD_TESTS) at standalone/CMakeLists.txt:245, and moqx's test/CMakeLists.txt:1 is find_package(GTest REQUIRED CONFIG). With BUILD_TESTS=OFF the tarball ships no lib/cmake/GTest/GTestConfig.cmake and that is a hard failure on any box without a CONFIG-mode system googletest. The comment at that gate anticipates it: CONFIG mode is used precisely because a module-mode-only system googletest will not satisfy moqx.

Nothing moxygen exports carries GTest — the only consumers are moxygen/test/ and moxygen/mlog/test/, neither installed, and moqtest_utils links only moxygen_moq. So there is no ABI reason for the two to share a googletest and moqx should bring its own. That is a moqx change, and BUILD_TESTS=OFF follows it.

Same shape as the samples ordering: two moqx changes gate a genuinely minimal artifact, one of them already in progress.

---

This change is 

[afrind]

There's a mix of things here. I think we want to do one PR that adds BUILD_MOQTEST and removes BUILD_TESTS, and leave BUILD_SAMPLES alone for now.

@afrind reviewed 9 files and all commit messages, and made 7 comments.
Reviewable status: all files reviewed, 5 unresolved discussions (waiting on gmarzot).

---

.github/workflows/omoq-version-release.yml line 246 at r1 (raw file):

            -DCMAKE_BUILD_TYPE=Release \
            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
            -DBUILD_TESTS=ON \

I believe we no longer need BUILD_TESTS either. The lone test util library consumed by moqx is now always built.

---

openmoq/scripts/collect-artifacts-standalone.sh line 57 at r1 (raw file):

# ── Step 0: Drop binaries that are not part of the relay artifact ────────────
# Built under BUILD_SAMPLES beside binaries we do ship, so excluded here.
rm -f "$INSTALL_PREFIX"/bin/moq_media_server* "$INSTALL_PREFIX"/bin/moq_mp4_receiver*

I think moq_media_server is still built under BUILD_SAMPLES, we can't turn off BUILD_SAMPLES until we convert all moqx tests to use moqtest instead of date/text, which is forthcoming.

---

standalone/CMakeLists.txt line 282 at r1 (raw file):

# Save original BUILD_TESTS/BUILD_SAMPLES/BUILD_TESTING values, then disable for fetched dependencies
set(_MOXYGEN_BUILD_TESTS ${BUILD_TESTS})
set(_MOXYGEN_BUILD_SAMPLES ${BUILD_SAMPLES})

This is premature

---

standalone/test/CMakeLists.txt line 3 at r1 (raw file):

# Regression guard for the per-layer XLOG category rooting in ../CMakeLists.txt:
# drives a short mvfst session and asserts each layer's --logging selector scopes
# to that layer. Skips (77) on bash < 4.

what is this?

---

standalone/test/local.cmake line 5 at r1 (raw file):

# xlog_category_pico is unregistered: its only drivers, pico_evb_relay_server and
# pico_evb_text_client, build under BUILD_SAMPLES, which nothing sets. Restore it
# once moqtest speaks picoquic.

? why would moqtest speak picoquic?

---

standalone/test/xlog_category_smoke.sh line 18 at r1 (raw file):

set -uo pipefail

# The marker maps and assert_scoping's nameref need bash 4; on 3.2 both fail

You can ask claude to make a bash3.2 compat version

[afrind]

@afrind reviewed 8 files and all commit messages, made 1 comment, and resolved 5 discussions.
Reviewable status: all files reviewed, 1 unresolved discussion (waiting on gmarzot).

---

openmoq/scripts/collect-artifacts-standalone.sh line 89 at r2 (raw file):

# whole folly/proxygen/mvfst stack, so a sidecar apiece adds ~130 MB of duplicate
# DWARF and pushes the -dbg tarball past GitHub's 2 GB asset limit.
NO_SIDECAR_BINS='moqchatclient moqdateserver moqflvreceiverclient moqflvstreamerclient moqtextclient'

moqx depends on moqtextclient

[afrind]

@afrind made 1 comment and resolved 1 discussion.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on gmarzot).

---

openmoq/scripts/collect-artifacts-standalone.sh line 89 at r2 (raw file):

Previously, afrind wrote…

moqx depends on moqtextclient

aha, this is just for dbg sections