feat: add YAML input support (-I yaml) with bundled libyaml 0.2.5

[vmvarela]

Closes #168

Summary

Add -I yaml as a supported input format. sql-pipe now reads YAML sequence-of-mappings as rows, mirroring the existing JSON input behavior (all-TEXT columns, fatal on parse error, exit code 2).

YAML is ubiquitous in CI/CD, configuration files, and infrastructure-as-code — this closes the biggest format gap in sql-pipe's input support.

Usage

# File auto-detected from .yaml extension
sql-pipe users.yaml 'SELECT * FROM users WHERE age > 18'

# .yml alias also works
sql-pipe users.yml 'SELECT * FROM users'

# Explicit -I yaml for stdin
cat data.yaml | sql-pipe -I yaml 'SELECT SUM(balance) FROM t'

# Works with all standard output formats and SQL features
cat data.yaml | sql-pipe -I yaml -O json 'SELECT name, amount FROM t ORDER BY amount DESC'

Changes

• lib/yaml/ (new, vendored) — libyaml 0.2.5 C sources and headers. Downloaded by CI from https://github.com/yaml/libyaml/archive/refs/tags/0.2.5.tar.gz; not tracked in git.
• build.zig — added addTranslateC for yaml.h (exposed as @import("yaml")); always-bundle 6 C source files (api.c, parser.c, scanner.c, reader.c, writer.c, loader.c). Skipped dumper.c/emitter.c (we don't emit YAML).
• src/format.zig — added yaml variant to InputFormat enum. Added explicit .yml → .yaml alias in fromExtension() (since std.meta.stringToEnum doesn't accept aliases).
• src/args.zig — help text for -I, --input-format now lists yaml.
• src/main.zig — .yaml arm in loadInput() dispatches to yaml_mod.loadYamlInput(). Imports the new module.
• src/yaml.zig (new, ~225 lines) — YAML loader using libyaml's event API (yaml_parser_parse + yaml_event_t). Mirrors src/json.zig:loadJsonArray: read-all, parse, all-TEXT table, insert rows. Every successful parse is paired with yaml_event_delete via defer.
• src/modes/{stats,schema,columns,validate}.zig — wired yaml_mod.loadYamlInput() so --columns -I yaml, --validate -I yaml, --schema -I yaml, and --stats -I yaml all work. (--sample correctly rejects non-CSV/TSV formats per existing design.)
• tests/fixtures/users.yaml + users.yml — block-style YAML mirror of products.json for testing auto-detect and the .yml alias.
• build.zig — 8 new integration tests (168a–168h).
• .github/workflows/ci.yml + .github/workflows/release.yml — added "Download libyaml amalgamation" step that fetches libyaml 0.2.5 to lib/yaml/ at build time (mirroring the existing SQLite download pattern).

Design Decisions

Decision	Choice	Rationale
Parser	libyaml 0.2.5 (C)	YAML spec is complex (anchors, flow vs block, multi-line strings, etc.); bundling a tested C parser avoids a 200-line minimal Zig parser that's guaranteed to fail on edge cases. Tiny footprint (~100KB source).
Always-bundle	No --bundle-yaml option	Unlike SQLite (which has version-sensitive features and is commonly system-installed), libyaml is tiny, rarely system-installed, and has no version-dependent features we need. System-library path would be untested dead code. YAGNI.
Event API vs document API	Event API (yaml_event_t)	Document API heap-allocates a yaml_document_t and adds complexity. Event API is sequential and gives per-event start_mark/end_mark for error reporting.
Structure	List of mappings only (flat)	Matches JSON's [{}] model. Nested mappings/sequences are rejected with a clear fatal error. Spec compliance trade-off: the original issue mentioned "flatten like JSON" for nested structures, but the plan (user-approved) restricted to flat mappings for v1. See "Follow-ups" below.
Type inference	All TEXT (no inference)	Matches JSON behavior. SQL can CAST(... AS INTEGER) as needed.
Anchors/aliases	Rejected with fatal	The event API doesn't resolve aliases; a no-op would silently corrupt data. Honest rejection is better than silent data loss.
Multi-document	Rejected with fatal	Silent data loss vs. schema mismatch from a second document. Error loudly.
.yml extension	Handled explicitly in fromExtension()	std.meta.stringToEnum only matches exact names. Without an explicit alias, .yml files would default to CSV.
Error format	"YAML parse error at line N, col M: <libyaml problem>"	Follows src/xml.zig:276-280 pattern. libyaml marks are 0-indexed; add 1 for user-facing.

Test Coverage (8 new integration tests, 168a–168h)

Test	Coverage
168a	.yaml file argument auto-detect
168b	.yml extension alias
168c	Stdin with -I yaml
168d	Filter + aggregate
168e	Comments transparent
168f	Malformed YAML → exit 2 with line number
168g	Empty input → graceful
168h	Multi-document → error exit 2

Validation

• zig build — clean
• zig build test — exit 0 (136+ tests: 128 existing + 8 new)
• zig build unit-test — exit 0
• ziglint src build.zig — exit 0, zero warnings
• Mode smoke tests — --columns, --validate, --schema, --stats all work with -I yaml

Memory Safety

Oracle-verified clean. Every yaml_parser_parse success path is paired with yaml_event_delete via defer. yaml_parser_delete is called exactly once via defer. All allocator.dupe results are tracked in a single defer block. sqlite_static ordering is safe because value frees happen AFTER sqlite3_step. On fatal exit, the OS reclaims all resources.

Follow-ups (not in this PR)

• C2: Nested structure flattening — The original issue asked for "support nested YAML structures (flatten like JSON does)". Current implementation rejects nested mappings/sequences with a fatal error. Follow-up issue will JSON-serialize nested values (like src/json.zig:114-121 does) to match JSON behavior.
• S1: Mode dispatch consolidation — Each of the 5 mode files (stats, schema, columns, validate, sample) has its own copy of the input format switch. Adding a 7th format requires touching 5 files. Pre-existing pattern, not introduced by this PR; would benefit from a future refactor.

Checklist

• Tests added (8 integration tests)
• Help text updated
• Build/lint/test all pass
• Memory safety reviewed
• CI downloads libyaml (mirrors existing SQLite pattern)
• No new dependencies for users
• No output format changes (input only, as specified)
• Branch name: issue-168/yaml-input (per AGENTS.md conventions)