fix(ci): preserve CUDA guard polarity in DSR ratchet

[localai-org-maint-bot]

Outcome

Fixes a fail-open branch-polarity defect in the CPU-only device-specific-reference (DSR) ratchet under ROAD-V1-A6 / BACKEND-SEAM-AUDIT S1.

The checker previously stored only whether a preprocessor condition mentioned a CUDA-family macro. As a result, #ifndef VLLM_CPP_CUDA and #if !defined(VLLM_CPP_CUDA) incorrectly exempted CUDA includes in portable code, while their CUDA #else arms were classified in reverse.

Design selected

I considered three approaches:

1. special-case only #ifndef, which leaves equivalent !defined expressions broken;
2. assign precedence to textual positive/negative terms, which fails on disjunctions;
3. model branch implication for CUDA-family-only boolean expressions, with conservative handling for unknown mixed expressions.

This PR implements option 3. CUDA-family predicates are evaluated under the portable-build assumption that all such macros are undefined. A false condition marks the current arm CUDA-only; a true condition marks its #else CUDA-only. Unknown disjunctions are counted conservatively. The accepted S1 contract and checkpoint records are updated without changing lifecycle, support, or the DSR baseline.

TDD evidence

RED first:

• 4 direct mutations proved #ifndef and !defined arms were exactly inverted
• 2 compound mutations caught a regression in defined(CUDA) && !defined(VT_OPTION)
• 2 disjunction mutations caught the CPU-reachable !defined(CUDA) || defined(VT_OPTION) arm

GREEN:

• focused polarity mutations 8/8
• full device-leakage mutation suite 32/32
• full script suite 103/103
• real tree remains DSR 32/32 with zero exemptions

Verification

• python3 -m unittest discover -s tests/scripts -p 'test_*.py' (103 tests)
• python3 scripts/check-device-leakage.py (DSR 32 equals baseline 32)
• python3 scripts/check-agent-record.py (ENGINE=131, MODEL=327, QUANT=82, KERNEL=45, BACKEND=68)
• python3 scripts/check-doc-checkpoint.py --commit HEAD
• README structure, model checklist, fusion consistency, and environment-documentation checkers
• git diff-tree --check HEAD
• independent code review: no Critical or Important findings

No runtime C/C++, baseline count, workflow, lifecycle, support claim, README, GPU, model, download, or performance behavior changes. Benchmark disposition is NOT APPLICABLE with benchmark_binding=false.

Slack

The required selection notification was attempted through the bundled secret-safe sender. No Slack channel is configured; #general returned channel_not_found. No credential was read or exposed.

FOLLOWING_AGENTS_PROTOCOL