Describe SparkCache behavior and qualification directly

[FujitsuPolycom]

Resulting documentation

The README begins with SparkCache as a persistent, rank-local NVMe context cache that reuses the longest verified stored prefix. State that cannot prove identity, compatibility, all-rank availability, and payload integrity is rejected, and vLLM computes the request normally.

The capability table distinguishes:

• tail-only row publication;
• tail-only opaque-page publication;
• different-root segment sharing;
• exact-prefix GPU reuse;
• SparkCache CUDA restore and placement;
• implemented behavior without live qualification;
• qualification limited to an exact artifact and deployment.

Artifact details are grouped under a dedicated qualification section instead of occupying the opening. The performance section highlights restore size and concurrency effects without reproducing benchmark methodology. Operations describe the default full-snapshot schema, opt-in copy-on-write tails, SSD-endurance limitations, and unsupported write-budget controls.

User-facing prose no longer calls the accelerated path native restore or native placement. Historical filenames and the sparkcache/native/ source directory remain unchanged because they are durable repository identifiers.

Compatibility

Documentation only. Cache identity values, digest salts, chunk geometry, stored schemas, runtime behavior, and the SparkCache CUDA placement ABI are unchanged. Cache namespace impact: none.

Validation

• SparkCache suite: 748 passed, 7 skipped
• Deployment suite: 108 passed, 1 skipped
• Ruff: passed
• Markdown conflict-marker, terminology, and whitespace checks: passed

This draft is stacked on the qualified SparkCache behavior consolidation in PR #32 and supersedes the earlier README draft in PR #24.

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 Closing this superseded draft and deleting only its remote head branch.