yaml: split module, raise conformance and performance

[davlgd]

This PR restructures vlib/yaml, raises conformance, and lands five measurable performance wins. No public API change.

Motivation

Current module pass 53/274 tests against the public yaml-test-suite. I wanted to raise conformance to better align with yaml lib from other languages and look at ways to enhance performance. It now passes 131/274 tests.

vlib/yaml.v had grown to ~1.9 KLOC in a single file mixing public surface, parser internals, flow grammar, tree traversal and emitters, so I've splitted it.

What changed

Module layout

vlib/yaml/yaml.v (1953 lines) is split by concern into 5 files (no API change):

File	Lines	Responsibility
yaml.v	401	Public types (Doc, Any, Null) and entry points (parse_*, decode/encode, Doc/Any methods)
parser.v	1042	Block-style Parser, scalars/quoted/comments, anchors/aliases, block-scalar chomping
flow.v	180	FlowParser for […] / {…}
path.v	139	Tree traversal (value_), dotted-key parsing, JSON↔YAML bridge
emit.v	202	YAML & JSON serializers

Conformance

• Tabs in indentation now report the offending line number.
• parse_quoted_string rejects unknown \x escapes instead of silently dropping the backslash.
• peek_next_indent and skip_ignorable consult the same directives_done flag, so %-prefixed lines stop being treated as directives once the body starts.
• collect_flow_continuation errors out on unterminated flow collections instead of returning silently truncated input.
• value_opt now distinguishes "absent key" from "key whose value is the explicit null literal".
• Decorator parsing (&anchor, *alias, !tag) extracted into a typed Decorators struct to avoid positional-binding bugs.

My yaml-test-suite runner test pass 131/131 on the targeted subset. It's intentionally not added to this PR — it requires a checkout of https://github.com/yaml/yaml-test-suite with many files. I've just added important tests of new supported cases.

What is not covered (deliberate, signaled by tests + comments): multi-document streams, explicit ? complex keys, custom tags, merge keys (<<:), full chomp/indent indicators (|2-), and a full anchors/aliases implementation.

Performance (representative ~1 KB doc, -prod -gc boehm)

	parse_text	to_json	to_yaml
Baseline (master)	16.2 MB/s	130 MB/s	11.9 MB/s
After this PR	22 MB/s	~350 MB/s	~30 MB/s
Gain	+36 %	+170 %	+150 %

Five separate commits, each independently measured:

1. to_yaml: stop calling json.encode per key — write directly into the active strings.Builder (+123 % to_yaml alone).
2. write_json_escaped_string: bulk-write safe runs via write_string instead of per-byte match (+170 % to_json on this fixture).
3. parse_scalar: skip to_lower() allocation behind a length-bounded ASCII case-insensitive compare; replace if contains_u8('_') { replace('_','') } with a single-pass strip_underscores (+30 % parse_text).
4. parse_quoted_string: return the body slice as-is when single-quoted strings have no '' and double-quoted strings have no \ — avoids the byte-by-byte rebuild in the common case.
5. JSON-superset fast path: parse JSON-shaped input directly into yaml.Any via parse_flow_value, skipping the json2.Any → from_json2 rebuild (+44 % on JSON-shaped input).

Further optimizations (lazy builder in gather_plain_continuation, 256-byte ASCII classification table for comment scanning) were prototyped and abandoned after A/B regressed — kept out of the PR.

Tests

• yaml_edge_cases_test.v (27 fns, 329 lines): parser edge cases with typed assertions
• yaml_conformance_test.v (19 cases, 145 lines): YAML-1.2 patterns from the spec
• yaml_json_roundtrip_test.v (30 cases + 500-iter idempotency, 82 lines): guards the JSON-superset fast path and the json2.Any rebuild path that previously crashed under -prod -gc boehm
• test_helpers.v: shared json_logically_eq (private, not _test.v so the helper can live in one place — V compiles each _test.v as its own binary)

All four test files green in both debug and -prod -gc boehm.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 604cd64fb9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Restore empty-document compatibility for typed decode

Returning root: null for empty/whitespace input changes yaml.decode[T] behavior for non-optional struct targets: Doc.decode now feeds "null" into json.decode, which errors, whereas the previous {} root let empty config files decode to default-initialized structs. This is a user-visible regression for callers that treat an empty YAML file as “no fields set”; please preserve the old decode-compatible shape or special-case Null before json.decode.

Useful? React with 👍 / 👎.

[davlgd]

I've fixed this and added a non-regression test