Consume hash-proven recurrent replay boundaries

[FujitsuPolycom]

Resulting behavior

SparkCache opts into vLLM recurrent replay-boundary hand-offs through supports_recurrent_boundary_blocks. A store plan is created only when SchedulerOutput.recurrent_boundary_blocks supplies exactly one (group_id, block_id, boundary_tokens) triple for every aligned recurrent group and every triple matches the request, store boundary, registered topology, and a non-null physical block.

The worker snapshots those proven blocks through the ordinary page-snapshot path. SparkCache never scans an arithmetic request-table slot or later running/speculative slots for a substitute.

Correctness

Absent, incomplete, malformed, wrong-boundary, wrong-group, null, duplicate, or preemption-stale metadata cancels publication. Request serving continues and recomputes on a later cache miss.

GPU-free regression D-17 models the live GLM geometry: a 6,912-token replay boundary, 2,304-token recurrent pages, a reclaimed arithmetic slot, and seven DFlash verification slots. It round-trips the explicit boundary through SparkContextCacheConnector and the real ManifestStore, then verifies byte-exact full-attention and recurrent restore. Negative coverage includes a stale non-null arithmetic slot and incomplete coverage across two recurrent groups.

The exact GLM vLLM lease contract accepts and requires the four recurrent-boundary runtime postimages.

Compatibility

Cache namespace impact: none. CacheIdentity values, digest salts, chunk geometry, manifest schemas, and page-delta wire bytes are unchanged.

Deployable SparkCache source SHA-256: 01cc59bf2c45af60f02813b771a484e65829e4f464eb27aa1756ca7067c82c9f.

This draft is stacked on SparkCache #29 at commit 19e2ec8b59c84ef359c2a3290f86962e3ff71d96. The paired runtime overlay is SparkRing draft #142 at commit 30b937be535db4113b4254458a2def1cf8e7535b. Integration order is this SparkCache consumer, then SparkRing #142.

Validation

• python -m pytest sparkcache -q — 752 passed, 7 skipped
• python -m pytest deploy -q — 108 passed, 1 skipped
• python -m ruff check . — passed
• Source-attestation tests — 2 passed

One concurrent validation run exceeded an existing sub-millisecond timing assertion. The isolated timing test and the following full SparkCache suite passed.

[FujitsuPolycom]

The resulting behavior and evidence are consolidated in retained draft stack #39 → #42 → #44. Independent heat and SSD research remains in #36. Closing this superseded draft and deleting only its remote head branch.