v0.1.2 - Performance pipeline and benchmark expansion

[0.1.2] — 2026-05-21 — Performance pipeline and benchmark expansion

This release adds the first performance gap-closure pass for benchmarked native
codegen and expands the local benchmark suite.

Codegen and runtime

• Added MIR-derived panic-effect facts so pure scalar statements and pure
  byte/list push values do not force unnecessary zmeu_panic_active() polls.
• Marked ordinary generated Zmeu helper functions private/internal and added
  inlining, symbol DCE, mem2reg, sroa, canonicalization, and CSE cleanup to
  the release MLIR pipeline.
• Removed impossible negative bounds checks for usize indexes and skipped
  proven redundant bounds checks for compiler-generated for loop element
  access.
• Added defended LLVM attributes for runtime declarations whose C contracts are
  already explicit, including nounwind and noreturn for process exit and
  fatal panic helpers.
• Changed list/byte-buffer initial growth from 1 to 8 and added a release
  runtime mode that disables the debug allocation registry while preserving
  overflow, null, allocator-tag, and allocation-failure checks.

Benchmarks and inspection

• Added benchmarks/ir_counts.sh for repeatable LLVM IR artifact counts.
• Expanded benchmarks/run.sh to discover benchmark fixture directories
  automatically.
• Added benchmark fixtures for branch-heavy scalar loops, helper calls, fixed
  array indexing, optional narrowing, borrowed range slicing, and struct field
  reads.
• Added docs/performance.md and docs/performance_plan.md as active
  performance tracking documents.

Validation

• Verified with python3 tools/gen_mlir_abi.py --check,
  cargo check --features mlir-backend, and
  cargo test --features mlir-backend.