param-op: an escaped ordinary character in a substitution pattern matches it (#750)

[berrym]

Closes #750 — but not the way the issue was filed. I filed it; the premise was wrong. See below.

The real defect (consensus-backed)

In ${var/pattern/replacement} the pattern is a pattern, so \X means "a literal X". The substitution path did not honor that for an ordinary character:

$ lush -c 'v=abc; printf "%s" "${v//\b/-}"'
abc                             # bash, zsh: a-c

lush_pattern_substitute chooses between the matcher and a plain substring search by looking for *, ?, [ or an extglob opener. \b has none, so it took the substring path and hunted for the two bytes \b. An escaped metacharacter (\*, \[) was fine, because the metacharacter itself satisfied the check.

Adding a backslash to that condition routes it to lush_pattern_match, which already implements the rule — case abc in a\bc) and ${v#\a} both work today. Dequoting here would have been a second copy of a rule with one canonical home.

This is the same defect [ was added to that condition for; the comment above it records that [bd] "never matched because the literal string never contained [bd]". One character class over, same shape.

The replacement half: curated, not fixed

The issue claimed ${path//X/\/} was the only way to get a literal /. Measurement says otherwise:

	lush	bash	zsh
${v//X//}	a/b	a/b	a/b
${v//X//usr/lib}	a/usr/libb	a/usr/libb	a/usr/libb
${v//X/\a}	a\ab	aab	a\ab

The delimiter is the first unescaped /, so every later / is already literal — a literal slash needs no escape at all, in one unquoted expression. And on the escape question bash and zsh diverge, with lush already following zsh.

So unescaping the replacement would have moved lush from zsh's model to bash's, with no capability gained and no rationale recorded — a curated default flipped by accident. Owner call taken: keep the literal rule, and document it so it stops reading as a bug (it is now pinned by tests, recorded in SEMANTICS §3.10, and explained above the splitter).

Also corrects a comment introduced with #684 that said bash and zsh drop the backslash. zsh keeps it — that error is what made #750 look like a defect.

Verification

• 14 new checks: escaped ordinary characters unanchored / replace-first / doubled / a space / both anchors / a non-occurring char; the three escaped metacharacters that must not regress; three pinning the curated replacement rule.
• 7 of the file's 40 fail against the parent build.
• Full suite 197/197. ASan 65/65; the test 1/1 under ASan.
• Every example in the new documentation was executed and matched as written.

One harness fix rolled in: check_agree now single-quotes its value. Unquoted, a b became a command prefix and a*b was globbed inside the array literal — which null_glob then emptied. Both looked like code defects and were purely the harness.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!