[dongfengy]

Description

Speculative overlap can advance device-side KV positions beyond host bookkeeping before draft KV is appended. Reserve an additional verification window, track the full allocated draft width through padding and unpaired draft growth, and reclaim unused slots without shrinking below non-helix committed history.

Use the same upper bound for runtime quotas, static cache sizing, and block-offset metadata. Also fail early when disaggregated one-model MTP through the legacy cache transceiver puts its draft layers outside the transferred target pool; connector-managed layouts keep their own validation.

Test Coverage

The existing disaggregated transition capacity case now expects 137 slots for a 128-token capacity and four draft tokens. Python AST parsing and git diff --check passed; independent source review checked runtime allocation against the common sizing helper.

No builds, tests, model execution, or benchmarks were run. Before leaving draft, validate overlap across block boundaries, padding/ragged speculation, dynamic draft reserves, disaggregated MTP, and helix-specific capacity accounting.

PR Checklist

• Description explains the behavior and scope.
• Commits include DCO sign-off.
• Runtime regression validation completed.
• Please check this after reviewing the above items as appropriate for this PR.