Vector quantization screen: all ArcadeDB and jvector schemes, Java layer and Python layer #586

Problem

I want matched-recall evidence for every quantization scheme (NONE, INT8, BINARY with rerank, PQ, jvector NVQ and BQ) on several dimensions, measured once in the Java layer and once through Python and SQL on the released wheels. Every round carries an A/A arm so the noise floor is known.

The screen decides which defaults and docs recommendations the vector index gets, and any engine proposal follows only from the pre-registered rule. The preparation on the released pair is done for one dataset, the timed cells are queued.

Priority: P2.

Acceptance

• Java-layer build and reopen cells are timed and scored for both datasets
• Layer-2 (Python and SQL) cells on the release pair are timed and scored for both datasets, with the A/A arm
• Per-metric verdicts at matched recall are posted, with churn and footprint covered by the follow-up issues
• Results are recorded in the study README before any recommendation is written

Results and the scorer output go in this issue as comments. Numbers are generated by script from the result files, never typed.