fix: ifFail dropping matchers, operator case, and quoted literals

[rapatao]

Summary

Four correctness defects in expression handling, a Kotlin engine speedup that
came out of fixing one of them, and a benchmark input fix that was pricing the
input's root kind as if it were input width.

Every defect below was reproduced first and has a regression test.

Bugs

ifFail discarded every matcher of a group. It rebuilt the expression from
left, operator and right only, so allMatch(...) ifFail FALSE returned a
fully empty expression, which then failed validation. It copies now.

A mixed case operator was unreachable by its own name. Operators are stored
under name().lowercase() but were looked up raw, so a custom operator named
myOperator could never be resolved. The lookup lowercases too.

A quoted literal spanning lines threw. The check was Regex("\".*\"") and
. does not match a newline, so the literal fell through to path resolution and
raised NoSuchElementException. The replacement is a startsWith and
endsWith check, which has neither the newline blind spot nor a regex on the
per-operand path.

Unknown operator: always printed null, because the message interpolated
the null local rather than the operator name.

Performance

Kotlin engine, three runs per configuration in one session, medians:

	ops/s	alloc per evaluation
before	2,225,259	849 B
after	2,598,127	375 B

An operand no longer runs toBigIntegerOrNull and toBigDecimalOrNull, the
second of which screens with a regex, unless it starts like a number, and paths
never do. Expression.isValid uses all rather than map plus firstOrNull,
which short circuits and stops allocating a list a group.

Benchmark input

The default input rooted at a data class while the wide variant rooted at a map,
so the two width columns differed by root kind as well as width. That is not
free: a map root is a hash lookup where a typed root is a reflective property
read, and it showed up as the Kotlin engine reading faster on the wider input.

Both root at a map now, and wideInput is built from narrowInput, so the two
cannot drift apart. BENCHMARKS.md is measured again throughout: all four tables,
the spread figures, the GC counts and the guidance.

Where the time goes is left alone. That measurement comes from a loop that is
not in this repository and cannot be reproduced from it.

CI

tests.yaml ran test on Java 17 against a Java 21 toolchain, which worked
only because the runner image happens to ship a 21. It runs check on 21 now,
which additionally enforces the detekt and 90% kover gates that were configured
but never actually executed.

Cleanup

• assertSingleCaseForDebugging, a debugging leftover pinned to case index 122
  that duplicated a parameterized case and would break if the case list shrank
• a BRANCH >= 90 kover bound declared twice
• unused credentialsVersion and jacocoVersion properties

Verification

./gradlew clean check passes, which covers the tests, detekt and the kover
thresholds for every module.