fix(cuda): allow missing runtime ELF build IDs

[rextio]

Root cause

A real TensorFlow 2.21.0 GPU/native run on WSL2 with an RTX 3060 reached successful native execution, numerical parity, and runtime-image mapping. Evidence production then failed only because tensorflow/libtensorflow_cc.so.2 has no GNU ELF build ID. The previous producer treated that optional ELF note as mandatory even though exact mapped canonical wheel paths, SHA-256 digests, byte sizes, roles, and mapped=true already bind the expected DSOs.

Change

• Record build_id: null when an expected mapped TensorFlow wheel DSO has no build ID.
• Keep lowercase bounded validation for every non-null build ID.
• Preserve exact role/path/mapping/SHA-256/size checks in the producer and offline verifier.
• Correct the manual evidence documentation and chain verification with &&, so a producer failure does not cause a secondary missing-evidence error.

TDD evidence

RED was observed before implementation:

• producer rejected mapped tensorflow_cc solely because build_id=None;
• verifier rejected null as a non-string.

GREEN verification:

• focused producer/verifier files: 54 passed;
• GPU-free suite: 414 passed, 1 skipped;
• Ruff: clean;
• mypy: no issues;
• python3.11 -m compileall -q src scripts tests: clean;
• documented shell chain: bash -n clean;
• git diff --check: clean.

Regression coverage also confirms malformed non-null IDs, path mismatches, unmapped and deleted images, wrong roles, bad hashes, and non-positive sizes remain rejected.

Residual validation

A fresh real WSL2 RTX 3060 manual harness run is still required to produce the trusted evidence artifact. This PR changes no support, certification, kernel-activity, transfer-profile, or performance claim.