test(concurrent): run concurrency.cpp tests at small size by default

[napetrov]

Part of #432

Summary

 // tests/svs/concurrent/concurrency.cpp
-#if TSan
-  small: 4000 + 1000 points, 50 queries
-#else
-  full:  20000 + 5000 points, 200 queries
+#if SVS_TEST_CONCURRENT_FULL_SIZE && !TSan
+  full:  20000 + 5000 points, 200 queries
+#else
+  small: 4000 + 1000 points, 50 queries
 #endif

New CMake option SVS_TEST_CONCURRENT_FULL_SIZE (default OFF, in tests/CMakeLists.txt) restores the full size for stress runs, e.g. a future nightly (#432).

Since #369, the four Concurrent MutableVamanaIndex stress tests make up more than half of serial ctest time, and they run in every x86/ARM/macOS matrix job, including ASAN Debug. Per-push CI cost went from 346 to 521 runner-minutes after #369 landed.

Evidence

Measured locally with taskset -c 0-3 (4 cores, the same as a GitHub-hosted runner), RelWithDebInfo, ctest -R "Concurrent MutableVamanaIndex":

Test	Before (full)	After (small)
batch iterator	214.6 s	42.5 s
search during mutation	131.6 s	24.7 s
quiescent recall	103.8 s	18.6 s
consolidate during search	96.5 s	19.5 s
6 tests total	~548 s	106.8 s

• All 6 tests pass. With --repeat until-fail:5, 30 of 30 runs pass, so the coarse recall thresholds hold at small size.
• The full serial suite drops by about 7 minutes per matrix job. The CI g++-11 log showed 1039 s for the full suite, 576 s of it in these four tests.
• -DSVS_TEST_CONCURRENT_FULL_SIZE=ON adds the define to the tests target only, which I checked in flags.make.

Merge Danger

Door: two-way

Blast Radius: tests-only

• Less stress per PR run: a race that needs a larger graph to show up would now only be caught by a full-size run, and no CI job turns the option on yet. Wiring it into a nightly is tracked in CI: tier the test matrix into PR / main / nightly to cut Actions minutes #432.
• No library or API change.

🤖 Generated with Claude Code