refactor: extract partial-eval constant phase

[lbliii]

Summary

• extract the literal-only evaluator, unresolved sentinel, expected-failure policy, and comparison dispatch into partial_eval_constants.py
• preserve partial_evaluate() and eliminate_dead_code() behavior through module-local compatibility aliases
• add focused tests for nested folding, dynamic fallback, short-circuiting, expected failures, invalid operators, and process-control exceptions
• reduce partial_eval.py from 2,307 to 2,174 lines

Why

This is the first bounded #143 phase split. Constant-only expression evaluation is a cohesive boundary shared by dead-code elimination and static-context evaluation, but it was embedded across the monolithic partial evaluator.

Impact

No public API, AST, render output, escaping, source-location, or optimization semantics change. The package smoke test confirms the new internal module ships in wheel and sdist artifacts.

Verification

• focused partial-eval suite: 319 passed
• make verify-stability: 4,459 passed, 5 skipped; coverage gate, safety suite, build, and clean-venv smoke passed
• repository-wide Ruff, format, and default-severity ty checks
• before/after: pytest benchmarks/test_benchmark_compile_pipeline.py benchmarks/test_benchmark_partial_eval.py --benchmark-only

Benchmark Evidence

• macOS 26.5.1 arm64, Python 3.14.2 free-threading build, gil_enabled=False
• 13 partial-eval render medians were flat or within about 4.2% of baseline; the worst movement was 2.000µs to 2.083µs and remains below the 5% regression threshold
• compile-pipeline medians improved 6–14% in this local run; treated as non-regression evidence, not a performance claim
• no committed baseline update; Linux CI remains authoritative

Steward Notes

• Consulted: Compiler/Codegen, Test Corpus, Benchmark, and root runtime guidance.
• Accepted: extract only the constant-expression phase and preserve both existing entrypoints.
• Required proof: direct phase tests, full partial-eval suite, full stability gate, package smoke, and before/after compiler benchmarks.
• No collateral: no public docs, examples, changelog, schemas, or migration notes because behavior and public contracts are unchanged.
• Deferred: control-flow, call/def inlining, loop-unrolling, and transform-phase extraction remain in Split partial evaluation into smaller phase-owned modules #143.

Refs #143.

[Copilot]

Pull request overview

Refactors the compiler’s partial-evaluation implementation by extracting the constant-only expression evaluation primitives into a dedicated module, keeping existing partial_evaluate() / dead-code elimination behavior wired through compatibility aliases while adding targeted regression tests.

Changes:

• Added partial_eval_constants.py to own the literal-only constant evaluator, the unresolved sentinel, expected-failure exception policy, and comparison dispatch.
• Updated partial_eval.py to import and alias the extracted constants/functions, removing the inlined implementations.
• Added focused tests covering nested folding, unresolved fallback, short-circuiting, expected failures, invalid operators, and process-control exception propagation.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
tests/test_partial_eval_constants.py	New focused test suite for the constant-expression partial-eval phase contracts.
src/kida/compiler/partial_eval.py	Rewires the monolithic partial evaluator to use extracted constant-phase primitives via module-local aliases.
src/kida/compiler/partial_eval_constants.py	New internal module encapsulating constant-only evaluation, sentinel, exception policy, and comparison dispatch.