Commit Graph

3 Commits

Author SHA1 Message Date
83e6e27d12 style(format): adopt --importgrouping testable-bottom from canonical rules-swift
Group @testable imports at the bottom, per public/rules-swift's swiftformat. A
behavior-preserving import-ordering reformat across 252 files (mostly tests).
--trailingclosures was deliberately not adopted (cosmetic, and swiftformat flags
it as caution-worthy).
2026-06-18 20:33:07 +02:00
78a8b628a5 feat(parsing): in-reader recovery keeps elements and pops to matching tags
Closes #93. Recovery moves into a recovering mode on EventReader, so malformed
input is repaired in place rather than dropped to the next markup boundary:

- A duplicate attribute is recorded and the later occurrence dropped, keeping the
  element. <a id='1' id='2'>y</a> now reads as <a id="1">y</a> with one
  diagnostic (was an empty tree with three).
- A mismatched or unexpected end tag pops the open stack to the matching ancestor,
  implicitly closing the elements nested inside it. <a><b><c>x</b></a> reads as
  <a><b><c>x</c></b></a> with a single diagnostic.
- Pre-root junk reports "unexpected character" instead of the misleading
  "content after the root element".

In recovering mode next() no longer throws: it records a Diagnostic and emits
best-effort events, popping to a match on a mismatch and stopping on truncation
so the builder closes what is open. The strict path is unchanged. readRecovering
becomes a plain non-throwing drain that takes the reader's diagnostics. Recovery
stays deterministic; new tests pin the kept-element and pop-to-match results.
2026-06-08 12:24:43 +02:00
b1d82e533d feat(parsing): recovering reader that never crashes on invalid input
Slice 2 of #92. Add PureXML.read, a deterministic recovering reader: it never
throws and never traps, returning the maximal best-effort tree plus one located
Diagnostic per problem.

- ReadResult { node, diagnostics } is the Equatable result type, so a
  deterministic read can be compared whole.
- Diagnostic carries a message and a structured source Mark; a thrown ParseError
  maps to one directly (ParseError.mark exposes the position each case carries).
- The reader resynchronizes after a failure (EventReader.recover(from:)):
  progress-aware, so a failure that already consumed the bad construct retries
  from there, while a stuck failure skips to the next markup start; it always
  advances when input remains, guaranteeing termination.
- The tree builder assembles leniently: elements left open by truncation are
  closed, an end tag with no open element is dropped. Well-formed input yields
  exactly parse(_:)'s tree with no diagnostics.

Tests pin exact trees and diagnostics for truncation, mismatched tags, and stray
junk, prove determinism (same bytes give an identical ReadResult), and fuzz 400
seeded inputs (arbitrary plus truncated valid docs) asserting no crash.

Recovery here is resync-based; finer in-reader recovery (keep the element on a
duplicate attribute, pop to a matching open tag) is the next slice.
2026-06-08 12:15:18 +02:00