CORE-9082

[ethan-wrasman-pkware]

Add documentation for known limitations, and fix a long-standing bug (originally introduced in 2014) where shorthand classes inside a character class produced invalid generations.

[a-z\d] was rewritten by a post-pass replaceAll to [a-z[0-9]]. Brics, the underlying regex engine, does not support nested character classes — it parsed the outer [...] as the class [a-z[0-9] followed by a literal ], so every generated string ended with one or more stray ] characters.

Shorthand expansion now happens inline while normalizing the pattern, with awareness of whether the cursor is inside a character class:

• Outside a class: \d → [0-9] (as before).
• Inside a class: \d → 0-9 (class-body form, no nested brackets).
• Negated shorthands (\D, \S, \W) inside a class expand to explicit complementary Unicode BMP ranges.

Also fixes a related off-by-one in [^X...] where the leading ^ was being emitted twice.

Added parameterized tests covering every shorthand in every position inside [...] (alone, with literal neighbors, with explicit ranges, in negated outer classes, under quantifiers), plus regression tests for every entry in the new LIMITATIONS.md.