Bind GLM-5.3 adaptive profile to tail publication

[FujitsuPolycom]

Result

The GLM-5.3 JJ 0b adaptive-MTP/fastsafetensors profile passes spark_cache_publication_schema=tail-cow-v1 in its serialized SparkCache connector config. The resolver rejects publication-schema drift before producing launch inputs.

Reason

SparkCache selects snapshot-v1 when spark_cache_publication_schema is absent. The executable profile explicitly selects the tail-only publication behavior specified by the quickstart.

Compatibility

The glm53-flash-hybrid profile uses block-page storage. SparkCache maps tail-cow-v1 to page-tail-cow-v1 in CacheIdentity and derives a tail-specific storage key. snapshot-v1 entries cleanly miss rather than being interpreted as page-tail objects.

The embedded-MTP digest identifies the target model, vLLM revision, maximum MTP depth, and adaptive policy. Publication schema occupies an independent CacheIdentity field, so the digest is 3255539158b8a4fd199b4d97d89eb5231df3b39a5370881c16099a8059b09e44.

The one-shot clear token is sparkring-b12x-kda-adaptive-mtp-fastsafetensors-initialization. Clear tokens are durable operator-action markers, not identity fields. Publication-schema selection does not require a root-wide clear because the tail-specific storage key prevents cross-schema restore.

Validation

• Focused profile contracts: 6 passed
• Ruff focused Python checks: passed
• Repository-relative Markdown links: 248 checked
• Maintained CPU suite with the origin/main baseline failure deselected: 1,919 passed, 9 skipped, 1 deselected
• Unfiltered maintained CPU suite: 1,919 passed, 9 skipped, 1 failed. The failure is scripts/test_glm53_flash_profile.py::test_public_glm53_benchmark_is_sanitized_and_bounded. Its assertion expects a README sentence absent from origin/main; the PR diff contains neither README.md nor that test.

No live or build host was contacted or modified.

[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 Closing this superseded draft and deleting only its remote head branch.