[stephanmoneybird]

Fix: #1125

The default branch of json_object_i converted the key with to_s without the mixed keys check the T_STRING and T_SYMBOL branches have, so { "1" => 1, 1 => 2 } generated {"1":1,"1":2} while { 1 => 1, "1" => 2 } raised. It now calls json_inspect_hash_with_mixed_keys when the first key was a String or Symbol. The TruffleRuby and Java generators already catch this case.

Hashes whose keys all have the same type pay one extra comparison per non String/Symbol key. A hash that starts with a String or Symbol key and then has another key type now runs JSON.on_mixed_keys_hash once, which is the cost the reverse order already paid:

== Encoding twitter.json
              before      4.586k (± 0.9%) i/s  (218.03 μs/i)
               after      4.521k (± 0.7%) i/s  (221.17 μs/i)
== Encoding citm_catalog.json
              before      2.233k (± 0.6%) i/s  (447.78 μs/i)
               after      2.204k (± 0.5%) i/s  (453.68 μs/i)
== Encoding integer keys (1000)
              before     18.825k (± 0.9%) i/s   (53.12 μs/i)
               after     18.522k (± 1.3%) i/s   (53.99 μs/i)
== Encoding small hashes with integer keys
              before      8.031k (± 0.9%) i/s  (124.52 μs/i)
               after      7.919k (± 0.9%) i/s  (126.28 μs/i)
== Encoding string key then integer keys (1000)
              before     18.751k (± 1.3%) i/s   (53.33 μs/i)
               after      6.049k (± 1.0%) i/s  (165.33 μs/i)
== Encoding small hashes with a string key then an integer key
              before     13.111k (± 1.0%) i/s   (76.27 μs/i)
               after      3.022k (± 0.6%) i/s  (330.86 μs/i)
== Encoding integer keys then a string key (1000)
              before      5.991k (± 1.5%) i/s  (166.90 μs/i)
               after      5.837k (± 1.6%) i/s  (171.33 μs/i)
== Encoding small hashes with an integer key then a string key
              before      2.980k (± 3.0%) i/s  (335.56 μs/i)
               after      3.048k (± 1.2%) i/s  (328.05 μs/i)

The integer keys cases are (1..1000).to_h { |i| [i, i] } and Array.new(1000) { |i| { i => i, i + 1 => i } }; the mixed cases put a "a" key before or after the Integer keys. Ruby 4.0.9, arm64-darwin.

So the mixed String-then-Integer order gets 3 to 4 times slower, the same as Integer-then-String already is. If that is too much, the cost is in JSON.on_mixed_keys_hash, which allocates a String per key and a Hash per inspected hash, and moving that check to C would speed up both orders.

A more general condition is possible as well: a single key_type != arg->first_key_type check before the switch, instead of the three per-branch checks, would also catch { 1 => :a, obj => :b } with obj.to_s == "1", which the C extension still generates while the Java and TruffleRuby generators raise. It would also inspect hashes mixing Integer and Float, or fixnums and bignums, which can never contain a duplicate. I kept the narrower check discussed in #1125, but I'm happy to switch.