perf: optimize identity function composition

[He-Pin]

Motivation:

bench/resources/cpp_suite/bench.07.jsonnet builds a lazy array of identity-equivalent composed functions. Current master evaluates the deep function/lazy chain normally; it is slow on JVM and still overflows Scala Native even with --max-stack 100000.

Key Design Decision

Recognize only statically safe identity-equivalent unary functions and keep the runtime check conservative. Direct identity (function(x) x) and the exact non-tailstrict self-composition shape (function(x) g(g(...g(x)))) can be elided only after the captured g is proven effectively identity. Recursive composition cycles are treated as non-identity so normal max-stack behavior is preserved.

Modification:

• Tag identity/self-composition function shapes in StaticOptimizer.
• Forward the tags from Evaluator into Val.Func instances.
• Add an iterative cached isEffectivelyIdentity probe with explicit Unknown/Yes/No/InProgress byte states.
• Short-circuit unary function application when the callee is effectively identity.
• Preserve laziness, tailstrict forcing, non-function errors, non-identity composition, and recursive-cycle behavior with evaluator tests.

Benchmark Results:

JMH on JDK 21, ./mill -j 1 bench.runRegressions bench/resources/cpp_suite/bench.07.jsonnet, lower ms/op is better; ops/ms = 1 / ms_per_op, higher is better.

Case	master ms/op	PR ms/op	master ops/ms	PR ops/ms	Delta
cpp_suite/bench.07.jsonnet	2.447	0.039	0.409	25.64	+6173% ops/ms (~62.7x faster)

Scala Native hyperfine, lower is better, comparing locally built Scala Native binary and source-built jrsonnet 0.5.0-pre98.

Target	Result
master native	StackOverflowError with --max-stack 100000
PR native	5.2 ± 1.1 ms
jrsonnet (-s 1000000)	28.4 ± 1.9 ms
PR vs jrsonnet	PR is ~5.46x faster

Analysis:

This optimization intentionally does not make arbitrary function equality assumptions. It only uses static shape tags plus runtime proof of the captured base function's effective identity. Errors and tailstrict are preserved: constructing f2(error ...) remains lazy, calling it forces the original error, and tailstrict still forces eagerly.

References:

• PR: perf: optimize identity function composition #826
• Rebased head: 50403de6fd8689a4cefb3cd39d280821ed33ff7b
• Base: 0ae7b78a93c4e643f9bcfb6dd1d99d9fe7a522a9

Result:

• ./mill --no-server -j 1 __.reformat && ./mill --no-server -j 1 __.test passed locally.
• ./mill --no-server -j 1 'sjsonnet.native[3.3.7]'.nativeLink passed locally.
• PR comments were reviewed; there are no review threads. The prior safety-update comment remains accurate after rebase.

[He-Pin]

Safety update pushed in 53556cfd:

• Added an IdentityChecking state so identity-composition probing detects recursive composition cycles and falls back to normal application instead of spinning in the optimization path.
• Added self-recursive and object-mutual-recursive regression tests.

Validation:

• ./mill -i sjsonnet.jvm[3.3.7].test.testOnly sjsonnet.EvaluatorTests
• ./mill -i sjsonnet.native[3.3.7].test.testOnly sjsonnet.EvaluatorTests
• ./mill -i bench.runRegressions bench/resources/cpp_suite/bench.07.jsonnet -> 0.039 ms/op
• ./mill -i __.checkFormat

[He-Pin]

Rebased onto latest master and refreshed local validation/benchmark data in the PR description. Master still overflows on the Scala Native bench.07 case; the rebased PR passes and remains strongly positive.