Consolidate SparkCache capability and prefix-reuse documentation

[FujitsuPolycom]

Resulting documentation

SparkCache documentation describes the persistent cache as it exists: longest all-rank stored-boundary discovery, authenticated rank-local objects, exact qualification boundaries, row-oriented prefix sharing, opaque-page limitations, and operator-visible storage behavior.

The qualification table retains one exact GLM-5.3 flat version 2 artifact:

• local image sha256:35b58a7bf414059c65b8f74e4e4b17ee6a81b7008e1bffbc9bd298b5e08c739e;
• SparkCache a1511d26a1fe2b17b24561bc52e376bf7f54b06a;
• source tree 4d5b8eb8c5c13793ee7a1e67b2b34bd38fcf4ddb;
• GLM-5.3 DFlash7 at TP4/DCP1 and C1;
• 131,072 restored tokens, 13 authenticated objects per rank, 1.35–1.50 seconds of sequential object read/authentication, 1.55–1.70 seconds of all-rank SparkCache CUDA restore, and the exact red codeword before and after restart.

No published OCI digest carries that source-bound runtime. The executable setup remains pinned to image 35b58a7….

Research limitations

Four-reader flat-page prefetch at SparkCache eabe7fd0c878db7384ef87fe80a1e96b9bedcf67 is research-only. Image sha256:df4e09a32cdbf1c0e69cc7c4c9e95d890d6c7a1e3eaac84f969912a16fd27dd3 structurally verified 813,068,464 bytes in 13 objects on all four ranks for a 131,072-token restore. It returned spark instead of the required red; a one-token-changed prompt with the same length recomputed red in 55.141 seconds. The candidate is not deployable and does not inherit the single-reader qualification.

The immutable record is evidence/glm53-flash-dflash7-bf16/flat-v2-four-reader-semantic-rejection-eabe7fd.json. Documentation and receipt tests require the research-only status, exact failed oracle, artifact identity, and non-deployable verdict.

Tail-only opaque-page deltas, shared opaque-page base reads, shared GPU-prefix attachment, and multi-root concurrent restore remain research-only because exact restored-output validation failed. Row-tail publication and different-root row-segment sharing remain implemented and GPU-free tested without live serving qualification. C16×128K remains unsupported with the recorded 20 GiB KV pool unless the runtime shares a GPU-resident trunk or provides additional KV capacity.

Compatibility

This documentation stack changes no cache identity, digest salt, logical chunk geometry, persisted schema, CUDA ABI, serving configuration, or package behavior. It does not merge, publish, close, delete, or operate a serving deployment.

Source identity

• Review base: SparkCache PR Research bounded flat-page object prefetch #43 head df1202686788995f174e410538e35c2978ebb526
• Documentation head: cf1a5820591fd8eea402c5f36c321e5def44ad3d
• Documentation tree: 49572ad51cf4539f6de32192f5dac19c6e9d433a
• Deployable source SHA-256 in both source-bound deployment profiles: 01643ea44598bf5ce2433b41768884ce975006488744a3d02a4600a7c05d6a3d

Validation

• python -m pytest sparkcache -q: 813 passed, 7 skipped
• python -m pytest deploy -q: 108 passed, 1 skipped
• python -m ruff check sparkcache deploy: passed
• HTML structure, local links, scenario labels, capability status, semantic-rejection wording, and immutable-receipt assertions: passed
• git diff --check: passed

The PR remains draft.

Review stack

PR #39 → PR #40 → PR #42 → PR #43 → PR #41. PR #43 is a research branch; PR #41 retains the qualified single-reader runtime in its public setup and support tables.

[FujitsuPolycom]

The explainer contract, flat-v2 C1 artifact evidence, four-reader rejection evidence, and present-state documentation are now retained in draft PR #44 through merge commit 7315377. Closing this absorbed draft and deleting only its remote head branch.