Read authenticated page-delta chunks concurrently

[FujitsuPolycom]

Resulting behavior

SparkCache reads and verifies page-delta chunk descriptors with a bounded eight-worker pool while preserving manifest descriptor order. Flat and delta-backed snapshots now use the same bounded concurrency model.

Technical reason

A qualified 17,152-token GLM page-delta restore spent 579.259 ms in storage reads because 67 authenticated delta chunks were read sequentially. The flat-snapshot path already uses eight workers.

Compatibility

Cache bytes, schemas, digest salts, chunk geometry, and namespace are unchanged. Existing entries remain compatible.

Validation

• SparkCache: 748 passed, 7 skipped
• Deployment: 108 passed, 1 skipped
• Ruff: passed
• Regression proves overlapping reads and deterministic descriptor ordering
• Deployable source SHA-256: bc238f96e550c7ec27d4081dd1f2e741d404aaf5c8572d89ccc5e76812be4d63

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 Closing this superseded draft and deleting only its remote head branch.