FIX InexactError when using allownan with typemax(Int)

[Kiwy3]

Summary

• correct parsenumber function and check_special macro to fix the bug
• add 2 tests (for Int64 and Int128) to ensure future versions will work with this

Root Cause

When using allownan, any number was parse as Float, and then convert as Int if needed. But the parsing as Float64 was rounded from 2e63 -1 to 2e63.

Notice: it's my first PR, so I may have done some mistakes. I ran all tests locally, and it was working.

Fixes #478

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.31%. Comparing base (233d644) to head (527057f).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #479      +/-   ##
==========================================
+ Coverage   89.99%   90.31%   +0.32%     
==========================================
  Files           7        7              
  Lines        1519     1528       +9     
==========================================
+ Hits         1367     1380      +13     
+ Misses        152      148       -4     

☔ 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.

[quinnj]

Thanks for the clear writeup and for jumping on a first PR the same day.

Diagnosis

With allownan=true, parsenumber currently skips the Int64/BigInt lexer and always goes through Parsers.xparse2(Float64, ...). Float64(typemax(Int64)) rounds to 2^63, and converting that to Int64 throws InexactError. Same class of bug for any integer with |n| > 2^53 (silent wrong value) and for typemax(Int128) / typemax(UInt64).

Always running the integer lexer, and only taking the float path for actual floats / specials, is the right fix and the right layer.

@check_special rewrite

This part I don't want to take as-is.

Master already double-increments i, so multi-char tokens almost never match — NaN / Inf / Infinity actually work via the Parsers fallback. The new loop matches correctly, but a prefix of the configured token at EOF still @goto invalid. Default inf is "Infinity", so:

JSON.parse("Inf"; allownan=true)    # UnexpectedEOF after this PR
JSON.parse("[Inf]"; allownan=true)  # still works (`]` → failed match → Parsers)

That's an inconsistency, and CI doesn't see it because the existing test overrides inf="Inf". Docs advertise allownan as allowing NaN, Inf, -Inf.

Also a behavior change worth calling out: untyped JSON.parse("123"; allownan=true) (and +1) will now return Int64 instead of Float64. That's the correct consistency with allownan=false.

Suggested fix

Either of these is fine:

1. Smaller patch: don't touch @check_special. Remove the if !opts.allownan wrapper, add elseif opts.allownan; isfloat = true, and change if isfloat || opts.allownan to if isfloat. Specials stay on the Parsers path like today.

2. Keep the new structure, but treat mid-token EOF as a failed match (then reset, which you already do) so "Inf" can fall through to Parsers as your comment intends.

Tests to add

• the original struct-field MWE from Inexact Error when parsing using typemax(Int) and allownan=true #478
• default "NaN" / "Infinity" / "-Infinity" and "Inf" / "-Inf" without overriding inf/ninf
• nested: JSON.parse("[Inf,NaN,-Infinity]"; allownan=true)
• typemin(Int64), 2^53+1, maybe typemax(UInt64)
• untyped JSON.parse(string(typemax(Int64)); allownan=true) === typemax(Int64)

Minor: trailing spaces on the new test lines, and "see note above" has no note.

Happy to merge once the Inf prefix case is fixed and those tests are in — this is a solid first PR.

[analysis & commentary by AI; reviewed by quinnj]

[Kiwy3]

Thanks for your fast feedback!

I reworked my PR without modifying the check_special macro, and added the mentioned tests, along with two additional tests to make sure the correct errors are thrown when using allownan.

I tried a few different approaches to handle all the specific cases covered by the tests, and I believe the current version should work correctly.

I also removed all the trailing spaces I had added previously, and I believe all the comments should now be clear and up to date.

If everything looks good to you, I'm happy with the current version. Otherwise, I'm happy to make any further changes if needed.

[quinnj]

[Reviewed by AI] Rechecked head 527057f. The original typed-integer failure is now covered by merged #480; current master 3e4cb97 passes the original struct case and Int64/UInt64/Int128 limit probes. This branch conflicts with current master in src/lazy.jl and test/parse.jl. Its full local suite also hits two signed-zero expectations that fail on the original base with the same dependencies. The new regression below is independent of those existing failures.

[quinnj]

[Reviewed by AI] P2: Keep configured numeric special tokens ahead of integer parsing. For example, JSON.parse("0"; allownan=true, nan="0") and JSON.parse("1"; allownan=true, inf="1") returned NaN/Inf on the original base; this branch returns integers 0/1 because the new integer path bypasses @check_special. Both baseline controls pass with matched dependencies. Current master already preserves these overrides through #480, so this logic should not replace that implementation.