[binggao1230]

Constant._pretty() writes every literal bare between single backticks, which breaks the pretty round trip (compile(pretty(g)) == compile(g)) on current main (8f9d741) in three ways:

import tatsu
tatsu.compile(tatsu.compile('start = ```a`b``` $ ;').pretty())    # FailedParse
tatsu.compile(tatsu.compile('start = ```a\nb``` $ ;').pretty())   # FailedToken

and silently, because the '' =literal '' alternative of the constant rule is tried before the raw /(.*?)/ path:

g = tatsu.compile("start = `'42'` $ ;")   # Constant(literal='42')
tatsu.compile(g.pretty())                 # Constant(literal=42) -- now an int

Same for 'True'/'None'/'true'/'4.5', and ' x' loses its leading space to the word rule. Alert._pretty() delegates to Constant, so alerts inherit all of this.

The fix picks a source form, from the constant syntax the grammar already accepts, that reads back to the same value:

• bare (output unchanged) whenever the value reads back as itself: plain words, spaced text, and all non-string literals;
• repr() of the value when one quote kind suffices — escaped quotes are not an option because the lazy string patterns stop at the first quote character, so repr's delimiter choice is what makes this parse;
• an '''...''' string, with quotes, backslashes, line breaks, tabs, and end whitespace escaped, when the value holds both quote kinds — multiline_string semantics apply trim(value.strip()).rstrip() before eval_escapes, so anything those touch has to be escaped.

Tests: a compile -> pretty -> compile harness asserting model equality (asjson) plus a pretty fixpoint, parametrized over 39 grammars covering the shapes above, and a check that already-stable literals keep their compact bare form. I ran the same round trip across the other node types (tokens, patterns, closures, joins, gathers, lookaheads, overrides, named, directives, comments): clean, except two pre-existing Pattern cases unrelated to this change (multiline patterns containing / fall back to ?"...", which cannot hold a raw newline, and multiline pattern text drifts to the trim() indentation fixpoint — the "weird indentation" note in grammar/tatsu.ebnf). Happy to file those separately.

uv run pytest tests: 641 passed, 6 skipped. ruff check --preview and ty check clean.

[codspeed]

Congrats! CodSpeed is installed 🎉

🆕 5 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• test_bench_parse_complex_expression: 325.1 ms
• test_bench_parse_arithmetic_expression: 124.5 ms
• test_bench_compile_calc_grammar: 46.5 ms
• test_bench_codegen_calc_grammar: 56.2 ms
• test_bench_compile_tatsu_grammar: 6.4 s

---