Push down Parquet scans with delete-aware ranges

[platypii]

Summary

• use hyparquet scan candidates for row-group, bloom-filter, and page-index pruning in both materialized and prepared reads
• retain absolute row positions so position deletes remain correct, and apply equality deletes before exposing deferred payload columns
• update to hyparquet 1.29.1 and reuse scan metadata/ranges instead of adding a separate callback path
• use a 128 KiB initial footer fetch for Parquet files below 8 MiB while retaining hyparquet's default for larger files

Closes #41.

Performance

Real HypAware cache

hypaware query overview --json --days 7 --include-local-only against a local cache containing 1,374 Parquet files. Implementations alternated each iteration, with two warmups per implementation and 11 measured runs:

	master	this PR	change
median runtime	2,206.0 ms	2,220.1 ms	+0.64%
output	6,462 bytes	6,462 bytes	identical SHA-256

A separate seven-run confirmation measured +1.12%. The non-delete many-file path is therefore effectively at parity, with a repeatable approximately 1% cost.

Selective reads and deletes

workload	master	this PR	result
one-file materialized read	1,077.8 ms	1,089.3 ms	1.1% slower
page-index selective read	0.463 ms	0.442 ms	4.8% faster
count with one position delete	33.45 ms	0.51 ms	65x faster
selective payload with one position delete	48.78 ms	5.74 ms	8.5x faster

The count-with-delete read volume fell from 2.491 MiB to 0.125 MiB. The selective payload query fell from 2.491 MiB to 0.388 MiB.

Footer range traffic

A read-only snapshot of 974 production Parquet files was evaluated using their actual file and footer lengths:

policy	bytes fetched	second requests
hyparquet 512 KiB default	48.1 MiB	0
adaptive 128/512 KiB policy	36.2 MiB	0

This is a 24.8% reduction in initial metadata traffic without another range request. The largest observed footer on a file below 8 MiB was 81.2 KiB. On the 160 raw AI gateway log files alone, the reduction was 22.5%.

Validation

• npm test — 59 files, 702 tests passed
• npm run lint
• npx tsc