Pin the live-tensor-compatible B12X runtime

[FujitsuPolycom]

Result

• Pins local-inference-lab/vllm@0b67266a0f37d6146a8403fb8482403c62f412d5 (tree ba9484ccb33aa56e90ff2f447f15ca9b9da97639) to local-inference-lab/b12x@b1d541f9e71a35f030d45fae437630fff7507c2a (tree c69cdec1c59a08e8e0e549f930fa8abcfb5134ae).
• Verifies the exact first-parent ancestry for both repositories and records it in the prepared-context receipt.
• Verifies that vLLM selects kda_metadata_validation="trusted" and passes request-sized live projection, metadata, and output tensors to bind_kda/run_kda.
• Verifies the installed B12X API and immutable commit/tree/contract labels before accepting an ARM64 image or resolved serving profile.
• Rejects B12X 2fcf23a0ce269be27b2e03fece73d46e90e6aeea, whose KDA API lacks trusted metadata selection and request-sized tensor binding.
• Runs the live-tensor runtime contract directory in the maintained CPU-only CI command.

Compatibility

Status remains implemented, not qualified. No runtime image was built and no serving deployment was inspected, stopped, or changed.

SparkCache wire fields, digest salts, 256-token storage geometry, and stored-object schemas are unchanged. The embedded-MTP draft identity schema now includes both the vLLM and B12X commits. Its digest is 2761488b43742f849e2cb7ac4977dc28d771faf5afef1951044327948d6ad71e, so entries written under another B12X implementation miss and recompute.

Validation

• python -m pytest runtime/glm53-flash-b12x-kda-adaptive-mtp/test_glm53_b12x_kda_adaptive_mtp_contract.py scripts/test_prepare_glm53_b12x_kda_adaptive_mtp_profile.py -q — 15 passed.
• Fresh prepare_context.py preparation followed by prepare_context.py --verify — passed for all pinned vLLM, B12X, NCCL, and InstantTensor inputs.
• ruff check --select E,F,W --ignore E501 on the changed Python files — passed.
• bash -n runtime/glm53-flash-b12x-kda-adaptive-mtp/build-image.sh — passed.
• The complete local offline suite reported 1,927 passed, 9 skipped, and 1 failure. GitHub CI, which now includes the live-tensor runtime directory, reported 1,928 passed, 9 skipped, and the same 1 failure. The sole failure is outside these files: scripts/test_glm53_flash_profile.py::test_public_glm53_benchmark_is_sanitized_and_bounded expects the sentence C4 and C8 were capacity-limited, which is absent from README.md at base revision 4e96afd03f1af79e6c7d2fbd6e8a7acdcb2b1d4b.

[FujitsuPolycom]

The resulting runtime, operator contract, and evidence are consolidated in retained draft stack #146 → #147 → #150. Independent page-base research remains in #149. Closing this superseded draft and deleting only its remote head branch.