perf: stack validated kube-prometheus optimizations

[He-Pin]

Motivation:
kube-prometheus and large_string_template remain the main source-built jrsonnet comparison targets for this stacked exploration branch. The accepted stack keeps only locally positive, output-equal changes and records rejected experiments so the same failed routes are not retried without a new hypothesis.

Key Design Decision:
Keep this PR as a benchmark-gated stacked exploration branch rather than a single small merge candidate. Each retained code change has local correctness smoke tests and benchmark evidence; higher-risk or noisy ideas are reverted and documented in bench/reports/sync-points.md.

Modification:

• Rebased He-Pin:perf/stacked-ready-gap-explore onto upstream master@1679892980e63e00945053eba02affca9dceae5f.
• Keep the committed kube-prometheus hot-path stack: std.manifestJson* visitor reuse, strict .json import fast path, strict JSON visitor trimming, strict JSON inline object layout, and strict JSON noOffsetPos reuse.
• Keep branch-local supporting work: ASCII-safe substr, ScopedExprTransform scope-map trimming, and single-pass static apply argument collection.
• Add the latest accepted large-template parser improvement: use String.indexOf('\n') for LF-only bulk ||| text-block line scanning while leaving CRLF/multi-character separators on the prior path.
• Update bench/reports/jrsonnet-gap-baseline-2026-05-10.md and bench/reports/sync-points.md with accepted and rejected benchmark evidence.

Benchmark Results:

kube-prometheus step	Scala Native result	Delta
Stacked baseline	235.971 +/- 12.925 ms	baseline
std.manifestJson* visitor reuse	224.975 +/- 11.550 ms	-4.66%
Strict .json import fast path	139.242 +/- 1.204 ms	-38.11% vs prior step
Strict JSON visitor trim	139.088 +/- 1.305 ms	-1.72% same-run A/B
Strict JSON inline object layout	136.301 +/- 1.957 ms	-2.60% same-run A/B
Strict JSON noOffsetPos reuse	135.9 +/- 1.2 ms	136.7 -> 135.9 ms on the final step

Large string template check	sjsonnet Scala Native	Source-built jrsonnet	Status
Post-rebase baseline	12.4 +/- 1.8 ms	3.3 +/- 0.6 ms	3.77x jrsonnet time
Accepted LF text-block scan, same-run	10.552 +/- 0.656 ms	5.611 +/- 0.826 ms	1.88x jrsonnet time
Candidate vs frozen clean sjsonnet	10.373 ms	clean 11.191 ms	-7.3% reverse-order Native A/B

Focused JMH / guard benchmark	Current result
large_string_template	0.683 ms/op
realistic2	41.618 ms/op
gen_big_object	0.822 ms/op
manifestJsonEx	0.052 ms/op
OptimizerBenchmark	0.414 ms/op after static-apply optimization
go_suite/substr	0.045-0.047 ms/op after ASCII-safe substr work

Analysis:
The strict JSON import stack remains the largest confirmed win, reducing the kube-prometheus run by roughly 42% from the original stacked baseline. After the latest rebase, large_string_template became the clearest string-heavy gap; the LF text-block parser optimization is narrow, semantics-preserving, and measurable, but the remaining source-built jrsonnet gap is still about 1.88x, so further parser/evaluator work is still needed.

Rejected routes now documented include Native base64 copy trimming, long-string escape-position collection, exact repeated-label format sizing, a specialized %(same_label)s scanner/render path, and lazy stdlib construction.

References:

• bench/reports/jrsonnet-gap-baseline-2026-05-10.md
• bench/reports/sync-points.md
• bench/resources/cpp_suite/large_string_template.jsonnet
• jrsonnet/docs/benchmarks.adoc

Result:
Branch head is 8f8ed59a2fe7ed8f2439c26049a60c9bef08b2c5. Local output equality, focused JMH guards, three-model review, and full ./mill --no-server -j 1 __.test passed for the latest accepted parser change.

[He-Pin]

I use this one as a references, will extract PR from it.