Align GLM-5.3 quickstarts with exact SparkCache contracts

[FujitsuPolycom]

Resulting behavior

The adaptive-MTP image builders, executable profiles, resolvers, and quickstart now bind SparkCache commit 5ec6a9953ad5d39120298bbfc26e95a6fa4b1dc3, tree 94c236b9dfbf5f70075eb47877fd9caaa5d8c249, and deployable-source SHA-256 bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63. This is the same SparkCache source contract used by the DFlash7 composition in the base branch. Resolvers require the complete canonical spark_cache_cuda_* configuration and reject incomplete or ambiguous configuration.

The DFlash7 quickstart selects the fastsafetensors target profile, includes strict preflight, start, rank-zero log tail, health, and generation-smoke commands, and explains the one-shot cache-removal token. A bounded evidence record binds the observed four-rank result to local image ID sha256:eef863d8bc578815a80b0e2d9f0d745102b6363415225101fd92171a2e5a55cb.

--prefill-schedule-interval 8 is recorded as a separate research-only mixed-traffic test. It is not added to either serving profile.

Status

• DFlash7 image construction: implemented.
• Exact local DFlash7 image: qualified only for recorded startup, health, continued generation, arbitrary page-boundary replay, and 131,072- and 262,144-token restore correctness.
• DFlash7 shared-prefix C2/C8/C16 on the exact image: unsupported by the retained record. Concurrency results from another image do not transfer.
• DFlash7 response quality and 262,144-token restore performance: unsupported and research-only, respectively.
• Adaptive MTP with SparkCache: implemented and GPU-free tested; four-rank persistent restore and performance remain unqualified.

Technical reason

The adaptive profiles used canonical SparkCache CUDA keys while their builders pinned a source revision that did not accept those names directly. Advancing the source contract makes the documented command path executable without an undocumented legacy-key rewrite.

Compatibility

The SparkCache pin advance does not change cache identity wire fields, digest salts, 256-token logical geometry, stored schemas, vLLM patch bytes, the lease contract, or the CUDA placement ABI. Compatible page-tail-cow-v1 entries remain eligible. The adaptive and DFlash profiles retain distinct model, draft, cache-root, and one-shot-clear identities. No public serving default changes.

Validation

• Maintained repository suite: 1,979 passed, 9 skipped.
• Focused runtime, resolver, quickstart, and evidence suite: 56 passed.
• Ruff over maintained Python trees: passed.
• JSON parsing and git diff --check: passed.

This draft is stacked on PR #143 and should be rebased or merged after that runtime contract is accepted.

[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.