Refuse JSON nested deeper than maxdepth instead of overflowing the stack

[quinnj]

Problem

Parsing recurses once per nested object or array. Input such as repeat("[", 100_000) (100 KB, well under typical request-size limits) throws StackOverflowError ("program state may be corrupted") from every entry point: untyped parse, typed parse(json, T), lazy(...)[], parse!, isvalidjson, and skipping unknown members. A server parsing untrusted bodies has no clean way to reject it.

Change

Every parse path enters a container through applyobject, applyarray, or the tuple maketuple. Each now checks a nesting depth carried on the LazyValue and, before recursing, throws:

ArgumentError: JSON nesting depth at byte position 513 exceeds the `maxdepth` limit of 512; pass a larger `maxdepth` keyword argument to parse deeper input

• New maxdepth keyword on lazy/parse/parse!/isvalidjson (default 512), passed like allownan or jsonlines.
• On master, the shallowest path (a recursive struct) overflowed at about 2,000 levels, so 512 keeps about 4x headroom.
• With error_context=true, the error arrives as the ParseError cause, like other errors.

Performance

• The depth and limit are Int32 fields in existing struct padding. LazyValue (80 bytes), LazyOptions (48), and LazyObject/LazyArray (72) keep their sizes.
• Allocations are byte-for-byte unchanged on all 56 read cases measured (the benchmarks/ cases plus larger realistic and container-only stress inputs).
• Instructions retired per call, measured against master: +0.4% to +2% on ordinary documents, and up to +5–6% on inputs made only of empty containers. The check is 5 instructions per object or array entered.
• Cycle counts and wall-clock time showed no difference outside measurement noise.
• The same 42 methods compile at first call, and the package image grows 0.3%.

Tests

• New nesting deeper than maxdepth testset. It covers 14 entry points/targets, the 512/513 boundary, raising and lowering maxdepth, and error_context.
• The trim entrypoint workload now compiles the limit error path and the maxdepth keyword under --trim.
• Two struct-target cases added to the existing truncated-input testset.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.60%. Comparing base (4f3dfdc) to head (e04da60).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #498      +/-   ##
==========================================
+ Coverage   92.57%   92.60%   +0.03%     
==========================================
  Files           7        7              
  Lines        1927     1935       +8     
==========================================
+ Hits         1784     1792       +8     
  Misses        143      143              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.