Four pre-existing CUDA ctest failures on main, attributed but unowned: test_capi (SIGSEGV), test_qwen3_apc_e2e, test_minicpm3_paged_engine, test_llama_paged_engine

[localai-bot]

Four CUDA ctest failures reproduce identically on unmodified origin/main —
same assertion, same line, same pass/fail counts — measured by a direct A/B on the
same box and toolchain. They are pre-existing and unowned. .agents/specs/triton-aot-default-on.md
records each as "owed its own issue"; none was filed. This is that record.

They are grouped here because they were attributed by one measurement pass, not
because they share a cause. Each needs its own investigation.

test	failure
test_capi	SIGSEGV at tests/capi/test_capi.cpp:480 — the ABI v8 custom-logits-processor case on a synthetic engine (MakeSyntheticEngine(), :243-248). 47/47 assertions passed before the crash; 3 of 4 cases green.
test_qwen3_apc_e2e	REQUIRE(anchor_ok) at :375, 59 assertions / 58 passed. Automatic prefix caching.
test_minicpm3_paged_engine	REQUIRE(first_div < 0) at :224, 26 assertions / 25 passed.
test_llama_paged_engine	attributed to main by the same arm.

Notes that will save the next person time

test_capi is not the documented flake. .agents/environment.md:189 calls
test_capi a -j-parallelism flake on this box, but this run was ctest -j 1,
so that explanation is unavailable and the SIGSEGV needs a real diagnosis. There is
prior art worth checking: a heap-corruption failure in test_capi was previously
seen on a different arch during sm_110 bring-up, so this may not be new.

Two of these are token-drift assertions against committed goldens
(test_qwen3_apc_e2e, test_minicpm3_paged_engine). Under "correctness first,
always" those deserve priority over the crash, since a silently wrong golden is
worse than a loud segfault.

Related, already tracked

• SACRED gate red on main: test_glm4_moe_lite_paged_engine diverges in tokens (plus 4 more pre-existing ctest failures) #233 — test_glm4_moe_lite_paged_engine (a SACRED gate red on main),
  test_linear_method, test_gemma4_registry_e2e, test_minimax_h3,
  test_serve_low_tools.
• test_qwen36_spec_decode: token drift in-suite, passes standalone 11/11 — filed as 'suite-context-dependent' but the decisive arm was never run #247 — test_qwen36_spec_decode, which is not in this set: it fails only
  in-suite and its attribution is still open.

Reproducing

Clean CUDA build on sm_121a, then ctest -j 1. Use -j 1 — ctest -j 4
OOM-rebooted this box (unified memory stacks concurrent model gates into host RAM).
Prefer in-suite runs: a standalone model gate here ran ~50x slower and reached
59 GB RSS on a 119 GB box, because the suite leaves the checkpoint warm in page
cache.

Found while gating #219 / #232; none is caused by that change.