Bound nesting depth to prevent stack-overflow DoS (GHSA-82x6-q7mm-w9cf)

[BinaryMuse]

Fixes security advisory GHSA-82x6-q7mm-w9cf: toml.parse() on deeply nested arrays or inline tables (a ~5 KB payload of a few thousand [ levels) drives the generated recursive-descent parser past Node's call-stack limit, crashing with an uncatchable RangeError. Since that's not a SyntaxError, application error handling that filters for parse errors rethrows it — a remote, unauthenticated DoS.

Fix

A depth guard on the value rule — the single point all array/inline-table recursion funnels through — caps nesting (default 500) and raises the library's normal parse error (.line/.column, via genError) once exceeded. The limit is configurable: toml.parse(input, { maxDepth }).

The counter is leak-free under PEG backtracking. The advisory's suggested ++depth … { depth-- } snippet decrements only on the success path, so every failed value probe (e.g. the one the parser backtracks over after a trailing comma) leaks an increment and can falsely reject valid input like a=[[1,],[2,],…]. value_choice here pairs the increment with a decrement on both the matching path (its action) and the non-matching path (a trailing predicate), so backtracking can't leak the counter — there's a regression test for exactly this.

The compiler recurses over the same AST but overflows later than the parser (survived depth ~2,200 while the parser crashed at ~2,500), so bounding parser depth also protects it.

maxDepth defaults to 500 — far beyond any realistic config nesting, well under the crash threshold, with headroom for smaller stacks (worker threads, reduced --stack-size). Setting it far higher re-exposes the overflow, so it's meant for lowering, not raising.