test(bb): run every root rollup check nightly, including the pinned vk

[AztecBot]

Summary

#24799 moved the heavy recursion suites to nightly but kept PinnedVKRootRollup per merge as the cheap change-detector for the root rollup circuit. The first merge-queue run after it landed shows it is not cheap.

From run 31409209377 (ci-full-no-test-cache, head c38ea15f, test-engine log 6b948df91b0fd544):

dsl_tests HonkRecursionConstraintTestWithoutPredicate/2.PinnedVKRootRollup   261s

That is the longest barretenberg test in the merge queue, against a default 600s per-test timeout — for a circuit whose full checks had already been deferred to nightly.

Comparing against the last pre-#24799 merge-queue run (31395115064, head 02217d37, test-engine log b96eb5be8d914778), the per-merge cost of the root rollup circuit went:

	before	after
…/2.GateCountRootRollup → …/2.PinnedVKRootRollup	29s	261s
…/2.GenerateVKFromConstraints	184s	nightly
…/2.Tampering	57s	nightly
total per merge	270s	261s

A wash — and with a longer critical path, since the work is now concentrated in one serial test rather than spread across three. The cost is constructing the ~6.35M-gate circuit at all, not the checks layered on top of it, so there is no cheaper per-merge detector to keep behind.

What this does

Drops the *.PinnedVKRootRollup carve-out in runs_nightly_only, so the whole of HonkRecursionConstraintTestWithoutPredicate/2 — pinned vk hash, pinned gate count, GenerateVKFromConstraints, Tampering — is emitted by test_cmds_nightly at CPUS=16:MEM=32g:TIMEOUT=60m.

Unchanged, so the smaller recursion circuits keep their per-merge detectors:

• the *.GateCount* carve-out — HonkRecursionConstraintTestWithPredicate/*.GateCountSingleHonkRecursion and ChonkRecursionConstraintTest.GateCountChonkRecursion still run per merge
• the debug-build carve-out for WithoutPredicate/1.GenerateVKFromConstraints, the only case exercising the debug-only native_verification_debug path

Also removes the HonkRecursionConstraintTestWithoutPredicate/2.* resource-prefix branch. Nightly tests get a fixed prefix before that code is reached, and no /2 case is emitted per merge any more, so the branch became unreachable; leaving it would have been a comment asserting a per-merge policy that no longer exists. The three comments that still described the pinned vk as the per-merge cover are updated.

The trade-off

The pinned gate count and vk hash become nightly-only, so a PR that changes the root rollup circuit merges green and the nightly goes red the next morning — the failure lands on next rather than on the PR that caused it. That is the deliberate cost of not paying 261s per merge for a circuit whose real checks already run nightly.

Worth knowing: the nightly job itself has never run. barretenberg-nightly-slow-tests.yml landed with #24799 at 14:33 UTC today and its cron is 0 6 * * *, so total_count for that workflow is still 0. It has workflow_dispatch — a manual trigger before this merges would validate the nightly emission end to end.

Testing

• Sourced nightly_only_tests + runs_nightly_only into a harness and classified the real test names taken from the two merge-queue logs above, under both clang20 and a debug preset:
  • all three WithoutPredicate/2 cases → nightly, in both presets
  • GateCountSingleHonkRecursion, GateCountChonkRecursion → per merge
  • WithoutPredicate/1.GenerateVKFromConstraints → per merge in debug only
  • AvmRecursiveTests.TranscriptOperations, IPARecursiveTests.*, ChonkTests.Basic → per merge, unaffected
• bash -n barretenberg/cpp/bootstrap.sh
• C++ changes are comment-only.

---

Created by claudebox · group: slackbot · requested by Sergei · Slack thread