OSH Parser

These facts are useful for the parsing contest.

15 lexer modes (lexical state)
233 IDs (token types / node types) in 23 kinds (core/id_kind_test.py shows this)
3 recursive descent parsers (command, word, [[)
1 Pratt parser (arithmetic)
[ fallback reuses osh/bool_parser
TODO: modify asdl.py to show these stats?
- X product types
- X sum types with X alternatives

what CPython opcodes does it use?
how many lines of code does it use in CPython? (Compare with execution.)
What is the distribution of ASDL string and array lengths per node type?
- note: there are several uses of string, not just token. Is this a good or bad optimization?

Brace detection -- this is a separate metaprogramming pass (doesn't depend on input). This is a recursive parser, although it operates entirely on token types and not chars/strings?
Per-Word Algorithms
- core/glob_.py
  - LooksLikeGlob
  - GlobEscape
  - GlobUnescape (in case of no matches, may not be necessary)
- regex escape, for passing to regcomp() (not done yet)
checking validity of names:
- for invalid-var in a b; do ...
- readonly invalid-var

core/word_eval.py -- after evaluating VarOp arguments, we compile globs to Python regexes, e.g. for ${x%foo*}
IFS splitting (this is quite slow and needs to be sped up!)
core/args.py -- this is not a recursive parser
echo -e -- backslash escapes (and printf if it turns out we need it as a builtin)
read without -r -- backslash escapes are parsed

Provide feedback