perf: preserve ASCII-safe simple format results

[He-Pin]

Motivation:
large_string_template still spent time re-encoding and re-scanning the huge string produced by simple named format interpolation, even when the static format literals and dynamic values made the final string JSON-string ASCII-safe. This keeps the focus on the largest remaining local gap from #666-style benchmarking.

Key Design Decision
Track ASCII safety as metadata on the compiled simple named format path instead of adding another renderer scan. The optimization is deliberately limited to all-simple %(key)s formats, where every emitted dynamic value goes through simpleStringValue and can be conservatively classified.

Modification:

• Add staticAsciiSafe metadata to compiled format strings.
• Return Val.Str.asciiSafe from Format.PartialApplyFmt when static literals and all simple named dynamic values are JSON-string ASCII-safe.
• Keep unsafe strings, unsafe static literals, and mixed-key unsafe values conservative.
• Add focused regression tests for safe numeric values, unsafe dynamic strings, unsafe static literals, and mixed-key safety.
• Update the gap and sync ledgers.

Benchmark Results:
JVM JMH (bench.runRegressions bench/resources/cpp_suite/large_string_template.jsonnet):

• Before: 0.683 ms/op
• After: 0.677 ms/op

Scala Native hyperfine, large_string_template, before vs after:

• Forward: 8.64 +/- 0.75 ms -> 8.01 +/- 0.52 ms
• Reverse: 8.65 +/- 0.50 ms -> 8.17 +/- 0.48 ms

Scala Native hyperfine, source-built jrsonnet comparison:

• sjsonnet after: 8.01-8.17 ms
• jrsonnet: 6.0 +/- 1.2 ms
• Remaining gap: about 1.34x

Guard benchmark, kube-prometheus:

• Forward: 131.74 ms -> 129.14 ms
• Reverse: 129.11 ms baseline vs 130.94 ms candidate
• Interpreted as neutral/noisy, not a target regression.

Analysis:
The existing renderer already has an ASCII-safe direct byte path, but formatted strings lost that metadata and fell back to UTF-8 encoding plus escape scanning. This change preserves the metadata at the producer where the safety condition is known, avoiding an extra whole-string encoding/scan on large simple format outputs. The safety predicate is conservative: unknown complex values, unsafe strings, or unsafe literal text do not get the fast-path marker.

References:

• performance optimization #666
• bench/reports/sjsonnet-vs-jrsonnet-gaps.md
• bench/reports/sync-points.md

Result:
large_string_template improves in both Native command orders, JVM JMH does not regress, output equality holds for large_string_template and kube-prometheus, and ./mill --no-server --ticker false --color false -j 1 __.test plus __.checkFormat pass.

Follow-up note:

• Rejected parser _asciiSafe hint reuse after PR creation: it improved debug parse_time in one run but regressed Native whole-process large_string_template in both command orders (8.23ms baseline vs 9.47ms candidate forward; 8.00ms baseline vs 8.47ms candidate reverse). The code was removed and only the ledger entry remains.

Follow-up note:

• Rejected Native manual ASCII-safe byte copy after PR creation: replacing String.getBytes(0, len, dst, dstPos) with a manual charAt loop preserved output but regressed large_string_template in both command orders (7.89ms baseline vs 10.78ms candidate forward; 7.97ms baseline vs 11.17ms candidate reverse). The code was removed and only the ledger entry remains.

Follow-up note:

• Rejected single-character StringBuilder.append(Char) fast path after PR creation: output stayed identical but Native large_string_template regressed in both command orders (7.97ms baseline vs 8.65ms candidate forward; 7.97ms baseline vs 8.84ms candidate reverse). The code was removed and only the ledger entry remains.

Follow-up note:

• Rejected ByteRenderer minified object comma/empty-state specialization after PR creation: output equality held and kube improved weakly (130.32ms -> 128.31ms forward; 130.62ms -> 129.74ms reverse), but large_string_template regressed/noised negative in both command orders (10.24ms -> 10.45ms forward; 10.99ms -> 11.27ms reverse). The code was removed and only the ledger entry remains.

Follow-up note:

• Rejected Native-only long ASCII escaped string renderer after PR creation: output equality held for kube, large_string_template, and a long Unicode fallback guard, but Native large_string_template regressed in both command orders (9.96ms -> 11.77ms forward; 9.56ms -> 10.54ms reverse). The code was removed and only the ledger entry remains.

Follow-up note:

• Rejected inline small-stack cycle tracking after PR creation: shallow/deep recursive error behavior and normal outputs matched baseline, but kube was noise-level and Native large_string_template regressed in both command orders (11.59ms -> 12.32ms forward; 9.67ms -> 10.80ms reverse). The code was removed and only the ledger entry remains.

Follow-up note:

• Rejected ByteRenderer quoted-key cache after PR creation: output equality held for kube, large_string_template, escaped/unicode keys, and long-key fallback, but kube reverse A/B was not stable-positive. HashMap was only weakly positive forward, while direct-mapped and capped variants regressed/noised negative in reverse. The code was removed and only the ledger entry remains.