Feature: transparent gzip input (.csv.gz files and stdin)

[vmvarela]

Context

Compressed data is the norm outside toy examples: log archives, data dumps, and curl | gzip pipelines all ship .gz. Today users must decompress first:

zcat huge.csv.gz | sql-pipe 'SELECT ...'   # extra step, loses file-as-table ergonomics

Zig stdlib provides std.compress.flate/gzip — zero new dependencies.

Proposed behavior

• File argument ending in .gz (or starting with magic bytes 1f 8b) is decompressed transparently before the format loader.
• Stdin with gzip magic bytes is decompressed the same way (sniff first 2 bytes of the buffered stream).
• Format detection uses the inner extension: data.csv.gz → CSV.
• Table name derived from the inner name: data.csv.gz → table data.

sql-pipe data.csv.gz 'SELECT COUNT(*) FROM data'
curl -s https://example.com/dump.ndjson.gz | sql-pipe 'SELECT * FROM t'

Scope (v1)

• gzip only (magic 1f 8b + .gz suffix).
• Streaming decompress into the existing buffered reader — no full in-memory inflation.

Excluded (YAGNI)

• zstd, xz, bzip2, zip archives
• Compressed output
• Multi-member gzip concatenation beyond what std.compress handles

Acceptance criteria

• sql-pipe data.csv.gz 'SELECT ...' works without zcat
• Gzipped stdin detected by magic bytes
• data.ndjson.gz loads as NDJSON, table data
• Clear error on corrupt gzip stream
• Non-gzip input path unchanged (no sniffing overhead beyond 2 bytes)
• Integration test with a gzipped fixture