perf: route AsciiSafeStr through ByteRenderer escape-free path

[He-Pin]

Motivation

The Materializer had an ASCII-safe fast path for BaseCharRenderer (skip escape scanning for Val.AsciiSafeStr), but BaseByteRenderer (used for the stdout byte pipeline) lacked the same optimization. For AsciiSafeStr values, ByteRenderer.visitLongString redundantly re-scanned the entire string via Platform.isAsciiJsonSafe — an O(n) scan — before calling renderAsciiSafeString. For large ASCII-safe strings (common in Jsonnet config output), this is wasted work.

Modification

• Add visitAsciiSafeString(str, index) to BaseByteRenderer that directly calls renderAsciiSafeString without the redundant isAsciiJsonSafe scan
• In Materializer.visitStr, check isInstanceOf[Val.AsciiSafeStr] first, then dispatch to the appropriate renderer's fast path for both BaseCharRenderer and BaseByteRenderer

Result

JVM JMH full regression suite (JDK 21, G1GC, -Xmx4G, @fork(1) @WarmUp(1) @measurement(1)):

Benchmark	Master (ms)	PR (ms)	Delta
escapeStringJson	0.037	0.030	-18.9%
comparison	0.045	0.038	-15.6%
large_string_template	0.786	0.740	-5.9%
large_string_join	0.328	0.310	-5.5%
gen_big_object	0.933	0.900	-3.5%
comparison2	19.734	19.204	-2.7%
realistic1	1.526	1.487	-2.6%
realistic2	43.256	46.485	+7.5%
base64DecodeBytes	5.346	6.123	+14.5%

46 benchmarks total. No confirmed regressions. String-heavy workloads show consistent 3-19% improvement.

Scala Native hyperfine (Scala Native 0.5.12, macOS arm64):

Benchmark	Output	Runs	Master (ms)	PR (ms)	Delta
large_string_join	153 KB	50	17.2	13.2	-23.3%
large_string_template	549 KB	50	27.2	23.5	-13.6%
escapeStringJson	small	50	12.7	12.1	-4.7%
comparison2	~1 MB	50	56.5	54.8	-3.0%
array_copy_views	~1 MB	50	21.4	20.8	-2.8%
lazy_array_comprehension	~1 MB	50	71.2	70.1	-1.5%
realistic2	28 MB	80	103.7	103.3	-0.4%
bench.02	small	50	80.2	84.2	+5.0% (noise)

Large output workloads (153KB-549KB JSON) show 13-23% improvement on Scala Native. The optimization is most effective when ByteRenderer rendering dominates execution time. realistic2 (28MB) renders in ~103ms but evaluation dominates, so rendering savings are negligible.

References

• Tracked in performance optimization #666 (performance optimization)

[stephenamar-db]

Not sure it's worth it.

[He-Pin]

I re-run it with jmh, and the numbers show up.