Harden pgvector 10M production evidence

[CaspianG]

Summary

• add bounded COPY ingest and halfvec storage for large pgvector runs
• validate remote row counts, index presence, checkpoint resume, and explicit prewarming
• add IVFFlat and binary-HNSW candidate profiles with original-vector reranking
• decouple durable ingest checkpoints from rebuildable index tuning

Verification

• 658 passed, 5 skipped before the final tuning additions
• 32 passed, 2 skipped on the focused streaming/workflow suite
• real pgvector 0.8.1 service smokes pass for halfvec, IVFFlat, binary HNSW, COPY, prewarm, and complete resume

Current evidence boundary

The 10M corpus and candidate indexes were exercised on a real Docker/WSL2 PostgreSQL service. The strict 10M result is intentionally not checked in yet: the monolithic profiles do not simultaneously meet recall >= 0.95 and p99 <= 100 ms. The next iteration is sharded pgvector fanout over smaller per-shard graphs.