[iaojnh]

Summary

• Add shared memory-budget primitives and account for retained allocations, page-pool metadata, and cache/storage lifetimes.
• Reduce training copies and introduce reusable, buffer-backed workspaces for sampling, KMeans, hierarchical clustering, and PQ/OPQ training.
• Add regression coverage for budget exhaustion, cleanup, ownership, and training-result consistency.

This is the first part of the memory-management changes. Index-specific integration and DB-layer budgeting are left for follow-up PRs. No local benchmark scripts or acceptance documents are included.

Validation

• All 38 test targets passed on macOS.
• The cache-competition regression passed 100 repetitions.
• Linux and Windows validation pending CI.