change!: fall back to the raw text when ansi-c unquoting fails

[Amey Pawar (ameyypawar)]

Created by Claude Code on behalf of Amey, who reviewed it before submitting. Everything below this line is the agent's writing, not his.

---

Summary

• ansi_c::undo() now fails when a quote is never closed, instead of returning the rest of the input as if it had been unquoted.
• Attributes and alternates fall back to the raw, still-quoted text when unquoting fails, which is what Git does at both of those call sites.
• For attributes this also covers invalid escapes, which previously discarded the whole line.
• The two Error::Unquote variants become unreachable and are removed.

Git baseline

unquote_c_style() in quote.c has no branch for running out of input: strcspn(quoted, "\"\\") stops at the terminating NUL, and the switch (*quoted++) that follows reaches default: goto error, returning -1. undo() instead treated a missing closing quote as success, so "abc unquoted to abc and reported all 4 bytes as consumed. A lone " already errored, so the two cases disagreed with each other.

Both callers in this workspace are places where Git recovers rather than aborts:

/* parse_attr_line(), attr.c */
if (*cp == '"' && !unquote_c_style(&pattern, name, &states)) {
	name = pattern.buf;
	namelen = pattern.len;
} else {
	namelen = strcspn(name, blank);
	states = name + namelen;
}

The alternates parser in odb.c says so outright — "Broken quoting (e.g., an entry that doesn't end with a quote) falls back to the unquoted case below". (That function is parse_alt_odb_entry() in 2.52.0 and parse_alternates() on master, so it is referred to by file here rather than by name.)

Measured with git check-attr, where the pattern is inferred from the file each line matches:

.gitattributes line	unquote	Git's pattern	matches
"abc	fails, unterminated	raw "abc	"abc
"\!hello"	fails, invalid escape	raw, and \! then escapes the !	"!hello"
"ab\qcd"	fails, invalid escape	raw	"abqcd"
"ab\tcd"	succeeds	ab<TAB>cd	ab<TAB>cd

The second row is why the invalid-escape case is included: gix-attributes turned it into an error for the whole line, while Git keeps the pattern and the leading ! never negates, because after the fallback the token still begins with a quote. Both implementations fall back before testing for the [attr] prefix, so "[attr]x stays a pattern on either side.

This does not extend to every unquoting site in Git — git checkout-index --stdin aborts with fatal: line is badly quoted on the same input — so the doc on undo() says callers differ rather than claiming a universal rule.

Validation

A differential sweep of 28 quoting shapes, taking Git's own answer via git check-attr rather than an expectation, goes from 12 divergences to 1. The harness was first run against an unmodified build to confirm it could detect the divergences at all, and corrected once when it scored dropped lines as agreement.

The remaining row is "\\", a pattern consisting of a single backslash. Both implementations produce that same pattern — the probe output is byte-identical before and after — and neither matches any file with it, so the oracle has no filename that can demonstrate agreement. It is a limit of matching-based comparison, not a divergence.

To show the refactor in parse_line() changes nothing it should not, 26 lines whose quoting succeeds or never begins were run through the old and new parsers: pattern text, mode, first_wildcard_pos, line number and every attribute assignment are byte-for-byte identical. The only behavioural difference is the failure path.

Reverting the gix-quote arm alone makes all three new tests fail.

• cargo test — gix-quote 17, gix-attributes 39, gix-odb 73, gix-pathspec 65, gix-dir 79, gix-worktree 10, gix-glob 40, gix 417
• cargo fmt --check — gix-quote, gix-attributes, gix-odb

gix-ignore is deliberately untouched: Git does not ansi-c unquote exclude patterns, so "quoted.txt" there ignores a file literally named "quoted.txt", and this crate already agrees.

Notes

The change is split as DEVELOPMENT.md asks, with the breaking change to gix-quote first and the two adaptations second.

It is confined to parsing. alternate::resolve() still applies its own handling to whatever path comes out, which is unchanged here.

One interaction worth flagging: #2847 hand-writes Display, Error::source and From impls for the Unquote variant in both gix-attributes and gix-odb, and carries a TODO(review) on the gix-odb one about it wrapping an Exn that does not implement std::error::Error. Removing the variants makes those impls, and that question, unnecessary. Whichever lands first, the other side is a deletion rather than a conflict of intent.

As for how the arm came to look like this: it was written before undo() reported consumed bytes, and a052d79 added that feature a couple of hours later in the same PR by extending this arm with consumed += input.len() — treating the unterminated remainder as consumed, rather than asking whether reaching the end should have succeeded at all.