[Perf] Reduce cold compile and cached restart latency

[alanray-tech]

Summary

• Store cached LLVM modules as bitcode, with backward-compatible textual IR loading.
• Use bitcode fingerprints for PTX caching and partition PTX by external optimization level.
• Cache immutable Python function-source preprocessing during repeated inlining.
• Skip dataclass AST flattening when a function has no struct locals.

Design notes

• Targets dev/compile-time-opt in this fork, so the existing graph/checkpoint stack stays outside this PR.
• Cache validation and specialization semantics are unchanged.
• Two speculative CSE optimizations were measured and rejected; they are not present in this branch.

Measured results

• Cold first step: 107.60 s -> 95.19 s (11.5% lower).
• Python materialization: 31.01 s -> 23.16 s (25.3% lower).
• Cached first step: 20.19 s -> 2.30 s (8.77x lower).
• Large LLVM artifact load: 16.83 s -> 0.84 s.

Test plan

• Offline-cache CPU/CUDA round trips: 2 passed.
• Offload and optimization tests: 35 passed.
• Kernel implementation, dataclass, and AST tests: 363 passed, 6 xfailed.
• Black, clang-format, whitespace, Ruff, and Pylint hooks passed.
• Genesis 200-step O1/O3 runtime guard completed with no overflow.

[alanray-tech]

AI attestation (2026-10-01): the code in this PR was generated by GPT-5.6 Sol. No manual code tweaks were made after generation. If a maintainer requests a stronger-model pass, I will switch to Claude Opus for the follow-up.

[github-actions]

Diff coverage: 100% · 10 lines, 0 missing