Feature: glob input (e.g. data/*.csv → single logical table)

[vmvarela]

Context

duckdb read_csv("logs/*.csv") treats multiple files matching a glob as a single logical table. sql-pipe today requires cat *.csv | sql-pipe for the same result. A glob flag eliminates the extra cat step and handles file discovery natively.

Proposed behavior

# All .csv files in a directory become one table "data"
sql-pipe "SELECT COUNT(*) FROM data" data/*.csv

# Mixed formats with -I override
sql-pipe -I ndjson "SELECT * FROM events" events/*.ndjson

Scope (v1)

• Use std.fs.openDirAbsolute + walk to resolve glob pattern into file list
• Concat file contents in-memory or stream sequentially into same table
• Validate that all files share the same header/columns
• Table named after the directory basename (e.g. data/*.csv → table data)
• Fatal error if any file has a different schema

Excluded (YAGNI)

• Merging files with divergent schemas (add missing columns, null-fill)
• Recursive glob (**/*.csv)
• Multiple glob patterns on the same command line
• File ordering control (sorted alphabetically by default)
• Union-all instead of append semantics

Files affected

• src/args.zig — detect glob characters (*, ?, [) in file paths, expand via filesystem walk
• src/loader.zig or new src/glob.zig — column validation across files, sequential loading
• src/main.zig — wire glob expansion into the existing per-file loading loop

Acceptance criteria

• sql-pipe "SELECT * FROM t" fixtures/*.csv loads all matching files into one table
• Schema mismatch across globbed files produces a clear fatal error
• Empty glob (no files match) produces a clear error
• Integration test using existing test fixtures directory