fix: lower TOML_MAX_NESTED_VALUES to prevent stack overflow on deeply nested arrays/inline tables

[danielbodorin]

What does this change do?

Lowers the default \TOML_MAX_NESTED_VALUES\ from 256 to 128 to prevent stack overflows when parsing deeply nested arrays ([[[[...) or inline tables ({a={b={c=...).

Each recursion cycle in \parse_array\ → \parse_value\ → \parse_array\ (and the equivalent \parse_inline_table\ path) consumes ~3,280 bytes of stack. At the previous default of 256:

• 256 × 3,280 = ~840KB — overflows 1MB stacks in sanitizer-instrumented builds (ASAN/MSAN)
• Observed crashes at ~191-310 recursion levels in OSS-Fuzz and Windows fuzzer builds

At the new default of 128:

• 128 × 3,280 = ~420KB — safe on 1MB stacks even under ASAN
• No real TOML document needs 128 levels of container nesting
• Users who need a different limit can still override via #define TOML_MAX_NESTED_VALUES \

PR #242 fixed a similar stack overflow for deeply nested dotted keys — this addresses the remaining array/inline table recursion path.

Fixes #292

Is it related to an existing bug report or feature request?

Yes — #292

Pre-merge checklist

• I've read CONTRIBUTING.md
• I've rebased my changes against the current HEAD of \origin/master\
• I've added new test cases to verify my change
• I've regenerated toml.hpp (how-to)
• I've updated any affected documentation
• I've rebuilt and run the tests with at least one of:
  • Clang 8 or higher
  • GCC 8 or higher
  • MSVC 19.20 (Visual Studio 2019) or higher
• I've added my name to the list of contributors in README.md

[marzer]

Thanks!

Though, please do not use AI to write issues and PR descriptions on any future contributions to this repository. I would much rather read something written by a human, than obvious AI prose filled with bad markdown syntax and unicode abuse.

[danielbodorin]

Thanks!

Though, please do not use AI to write issues and PR descriptions on any future contributions to this repository. I would much rather read something written by a human, than obvious AI prose filled with bad markdown syntax and unicode abuse.

Go it! Thanks!