test(ci): give ASAN recursion-constraint test a 20m timeout

[AztecBot]

Why merge-train/spartan-v5 (PR #24272) keeps getting dequeued

The train PR's merge-queue CI (merge-queue-heavy, 10 grind runs) fails because a single ASAN unit test times out, not because of anything in the train's commits.

The failing test, on every grind run:

FAILED: barretenberg/cpp/build-asan-fast/bin/dsl_tests \
  --gtest_filter=HonkRecursionConstraintTestWithoutPredicate/2.Tampering (601s) (code: 124)

code: 124 is a timeout kill, not an assertion failure. The dashboard history for this test on v5-next shows it creeping up over a few days — ~270s (06-24) → ~400s (06-25) → >600s (06-26), and now all 10 grind runs hit the 600s default and are killed, dequeuing the train on each Auto-merge/dequeue cycle.

Root cause

This ASAN test proves and verifies a full UltraHonk circuit. Its cost has grown with the verifier circuit on the v5 line. In barretenberg/cpp/bootstrap.sh, test_cmds_asan emits every ASAN test with CPUS=4:MEM=8g, so they inherit the 600s default timeout (ci3/source_test_params: TIMEOUT=${TIMEOUT:-600s}). The full suite runs through non-strict parallelize (CPU-based concurrency + GNU-parallel --memsuspend backpressure), so under the parallel grind the test slows well past 600s and is killed.

Fix

Give just this heavy ASAN recursion test a 20m timeout — the same budget already used for the other heavy recursion proving test (bbapi_tests ChonkPinnedIvcInputsTest.AllPinnedFlows, TIMEOUT=20m). One-line behavioral change: the emitted command becomes …:CPUS=4:MEM=8g:TIMEOUT=20m ….

Local verification

• Built build-asan-fast dsl_tests and reproduced the original grumpkin g1 data not found only because the warm CRS was the pre-v2 format; after barretenberg/scripts/download_bb_crs.sh (fetches grumpkin_g1_v2.flat.dat), the test runs for real.
• Ran the exact command: [ OK ] HonkRecursionConstraintTestWithoutPredicate/2.Tampering (178752 ms) — passes in ~3min in isolation (peak ~10.9 GB). So this is a wall-clock-budget problem under grind contention, not a logic failure or hang; a longer timeout is the correct, minimal fix.
• Confirmed source_test_params parses the multi-segment prefix and exec_test applies timeout -v 20m.

Full ./bootstrap.sh ci (merge-queue-heavy, EC2-orchestrated) can't run on this host, so ci-full is set here to exercise the ASAN path through PR CI.

This is shared CI infra; landing it via the spartan-v5 train carries it into v5-next, where the regression is observed.

---

Created by claudebox · group: slackbot

[AztecBot]

Automatically closing this stale claudebox draft PR (no updates for 5+ days). Re-open if still needed.