Optimize JsonCompleter.parse append-only byte hot path

[default-anton]

Summary

Optimize JsonCompleter.parse for append-only streaming input.

The old hot path still paid for input.start_with?(input_snapshot) on every growing chunk to prove the new input shared the old prefix. On long streams that turns into a repeated memcmp over already-processed data and showed up as the dominant profiler frame. This change keeps the existing parsing behavior for append-only streams, switches the syntax dispatch/scanners to raw ASCII bytes, and adds inline comments around the byte-oriented path so the optimization stays reviewable.

Before / after

Benchmark command:

JSON_COMPLETER_BENCHMARK=1 bundle exec rspec spec/parse_benchmark_spec.rb

Benchmark settings from the spec run:

• payload bytes: 77690
• prefixes: 9712
• iterations: 50
• chunk size: 8

parse

Metric	Before	After	Change
total runtime	1.2887s	0.8435s	34.5% faster
per iteration	25.774ms	16.870ms	34.5% faster
allocated objects	2,047,568	994,865	51.4% fewer
heap growth bytes	196,608	0	eliminated

Relative to complete + JSON.parse

Metric	Before	After
speedup	12.93x	19.89x
allocation reduction	8.83x	17.14x

What changed

• removed the repeated full-prefix start_with? check from the append-only parse hot path
• kept parse state resets for truncation, same-length edits, and completed top-level values
• switched parser/completion dispatch and shared scanners to byte-oriented loops (bytesize/getbyte/byteslice)
• added an incremental multibyte string parse spec to guard the byte path
• documented the append-only stateful-instance contract in README.md and CHANGELOG.md
• added inline comments explaining the byte constants and why the optimization is safe for UTF-8 payloads

Why this helps

Most streaming updates are append-only and mostly boring bytes.

Before:

• compare the whole prior prefix with start_with? on every chunk
• branch on 1-character strings in the main dispatch/scanner loops
• keep paying that cost as the stream grows

After:

• trust append-only stateful input and only reset on the cases that actually need it
• branch on ASCII byte values instead of allocating/transcoding tiny strings
• keep multibyte string content intact by copying plain runs with byteslice

Same result, less repeated work.

The profiler signal matched that story: rb_str_start_with dominated the sample before the change and disappeared from the hot-path sample after it.

Validation

• bundle exec rubocop
• bundle exec rspec
• JSON_COMPLETER_BENCHMARK=1 bundle exec rspec spec/parse_benchmark_spec.rb