fix(circuits): bench each protocol circuit under the scheme its VK uses

[AztecBot]

Problem

The rollup circuits have not been benched since the per-kind VK change in May. bench_cmds in noir-projects/fnd/noir-protocol-circuits/bootstrap.sh picks the bb gates scheme for each artifact by grepping its name against ivc_patterns. That array does not exist: the chonk patterns are read from chonk_circuits.json into ivc_kernel_patterns and ivc_app_patterns, and write_vk matches them through the split regexes, while bench_cmds kept the old name. An unset array expands to one empty line, and an empty pattern in a grep -f file matches every input, so the first branch always wins and every artifact is benched with --scheme chonk.

bb rejects that for the rollup circuits, and run_bench.sh pipes the failure into jq without pipefail, so the command exits 0 and writes an empty bench file that bench_merge then drops. In the last merge-queue bench run (engine log) every run_bench.sh ./target/rollup_*.json --scheme chonk line is PASSED (0s); the rollup_tx_merge sub-log is just Rollup Honk proof type not supported on MegaBuilder. The chonk circuits happen to land on the right scheme, so the kernel series were unaffected.

Surfaced while reviewing #25506, which relies on the bench to show reset-variant cost changes.

Change

• VK generation and the bench classify a circuit through one function. circuit_kind maps an artifact name to hiding, kernel, app, rollup_honk, rollup_root or ultra_honk, in the order write_vk already applied. write_vk and bench_cmds each turn the kind into their own bb flags: write_vk keeps --circuit_kind for the chonk kinds and --oracle_hash keccak for the root, neither of which bb gates takes. The combined ivc_regex export had no remaining caller and is removed.
• run_bench.sh runs under set -euo pipefail, so a bb error fails the bench command and, through the bench engine's fail-fast, the bench run. It also resolves bb through find-bb, the AVM-enabled binary the build keys the circuits with, instead of the hard-coded non-AVM build/bin/bb: rollup_tx_base_public verifies an AVM proof and the non-AVM binary refuses it.

Effect on the bench

The rollup circuits' _gates and _opcodes series reappear. The kernel, app and hiding series are unchanged.

Tests

• scripts/circuit_kind.test.sh pins the flavor each circuit family lands on, including the mock circuits' names, which share this script. It is wired into test_cmds, keyed on the bootstrap script, the test and the two pattern files it reads, so CI runs it under noir-protocol-circuits-tests. It exists because the build only runs the classifier for a VK that misses the cache: on a warm cache (example) every VK is downloaded and write_vk never executes, so a misclassification would otherwise surface on the next cold-cache build.
• Run locally against the published 6.0.0-nightly.20260902 artifacts with the release bb of the same version: every non-simulated artifact measured under the flags the fixed bench_cmds assigns it. 49 of 50 succeed; the one failure is rollup_tx_base_public under the non-AVM binary ("AVM recursion is not supported in this build"), which is what the find-bb change addresses. The bench-foundation target builds bb-cpp-native, which produces bb-avm alongside bb.
• run_bench.sh exercised directly: a correct invocation writes the two-entry bench file; a wrong-scheme invocation now exits 1.
• ./bootstrap.sh bench_cmds against placeholder artifacts covering every class plus a _simulated one: simulated skipped, chonk kinds get --scheme chonk, rollup honk gets --scheme ultra_honk --ipa_accumulation, root and the rest get --scheme ultra_honk.
• bash -n on both scripts.

---

Created by claudebox · group: slackbot · requested by Mike (@iAmMichaelConnor) · Slack thread