fix(bb): repair acir components debug regression test

[AztecBot]

Summary

• Fixes the canonical nightly debug failure in run 30190406234, where AcirComponentsCheckTest.DetectsUnconstrainedWitnesses aborted under _GLIBCXX_DEBUG with an out-of-bounds std::vector access.
• The root cause was the regression test corrupting UltraCircuitBuilder::real_variable_index with resize(9). That violates builder invariants, so the static analyzer aborted before ComponentsChecker could report the intended UNCONSTRAINED error.
• Reworks the fixture to compare the full ACIR circuit against a valid synthesized circuit that omits the ACIR witnesses, and adds the checker message to the expected-type assertion for easier future diagnosis.

Notes

The linked aztec-claude run 30190267425 is a separate stale scheduled workflow copy; aztec-claude#2041 already carries that schedule guard. The real current debug-build failure is the canonical aztec-packages run above.

Testing

• cmake --preset debug
• cmake --build --preset debug --target acir_components_check_tests -j$(nproc)
• ./build-debug/bin/acir_components_check_tests --gtest_filter=AcirComponentsCheckTest.DetectsUnconstrainedWitnesses
• ./build-debug/bin/acir_components_check_tests
• NATIVE_PRESET=debug barretenberg/cpp/scripts/run_test.sh acir_components_check_tests AcirComponentsCheckTest.DetectsUnconstrainedWitnesses

---

Created by claudebox · group: slackbot · Slack thread

[AztecBot]

Independently reproduced and verified this fix from tonight's nightly dispatch (aztec-claude run 30333107696). This PR is correct and ready to land — it just needs a human to mark it ready and merge.

Posting the verification because the PR did not yet carry test evidence, and because this is the blocker for a failure that has now recurred every night since 2026-07-16.

Verified locally against the nightly's exact preset

Built cmake --preset debug (which sets -D_GLIBCXX_DEBUG — the only preset that catches this) on next @ 63e8535ef93:

• Red (base next, unpatched): AcirComponentsCheckTest.DetectsUnconstrainedWitnesses aborts with exactly the CI message — attempt to subscript container with out-of-bounds index 10, but container only holds 9 elements.
• Green (this PR's components_check.test.cpp applied): acir_components_check_tests 10/10 pass.

Confirmed root cause via debugger

The stack confirms the diagnosis in the description — the abort is inside the analyzer, before the checker's own bounds guard:

components_check.test.cpp:566  checker.check()
components_check.cpp:31        build_circuit_component_map() -> StaticAnalyzer_ ctor
graph.cpp:394                  process_execution_trace()
graph.cpp:84                   extract_gate_variables()
graph.hpp:105                  to_real() -> real_variable_index[10]   // size 9

CircuitBuilderBase::add_variable appends to variables and real_variable_index in lockstep, so real_variable_index.size() == variables.size() is a builder invariant that nothing in production ever breaks. resize(9) broke it, and to_real maps every gate wire through that vector.

The replacement test still covers the intended guard

Worth stating explicitly, since a rewritten fixture can easily end up passing vacuously: I temporarily deleted the

if (witness >= builder_.real_variable_index.size()) { ... }

guard in components_check.cpp and re-ran this PR's version of the test — it fails (out-of-bounds index 1001, but container only holds 11 elements). So the fixture genuinely still exercises that branch.

No regressions

• Release preset (--preset default): acir_components_check_tests 10/10, boomerang_value_detection_tests 96/96.
• Debug preset: boomerang_value_detection_tests 96/96 (the main StaticAnalyzer_ consumer).

Why this keeps recurring

This is the 11th ClaudeBox branch opened for this same failure since 2026-07-16. They are opened as drafts, never reviewed, and auto-close-stale-drafts.yml closes them after 5 days — #24783 was closed that way on 2026-07-23 with no human ever looking at it. Each night then re-dispatches and a new duplicate branch appears. I deliberately did not open a 12th; merging this one breaks the loop.

Optional follow-up (not needed to land this)

To stop the failure mode recurring silently, a one-line invariant check in the StaticAnalyzer_ constructor turns a future opaque OOB abort into a clear message at zero release cost:

BB_ASSERT_EQ(circuit_builder.real_variable_index.size(),
             circuit_builder.get_num_variables(),
             "real_variable_index and variables must be the same length");

I verified this compiles and passes both suites in debug and release. Happy to push it here or open it separately — say the word.

---

Created by claudebox · group: slackbot