# Phase 125 completion evidence

Verified on 2026-07-13 in `/mnt/volume300/workspace/oshun-lt-2`.

## Native solver and bindings

- `cargo test --offline`: 10 passed.
- `cargo clippy --all-targets --features napi-bindings --offline -- -D warnings`:
  passed.
- `cargo build --release --features napi-bindings --offline`: passed and
  produced the napi-rs addon copied by `copy-native.mjs`.
- Native parity tests exercised every exported equilibrium kernel with 100
  deterministic random inputs and canonical MC with 100 deterministic parameter
  sets at 12-digit, physical-scale precision: 4 passed.
- The native addon is exposed through the `@kalika/thermodynamics/native`
  package subpath with deterministic TypeScript fallbacks for equilibrium
  kernels.

## Scientific benchmarks

- Fe-C: eutectic/eutectoid invariant temperatures, 0.76 wt% C eutectoid
  composition, and austenite carbon solubility checks pass.
- Cu-Zn: alpha solubility, beta ordering, and all five assessed peritectic
  reactions pass.
- Al-Cu-Mg at 200 °C: theta, S, and T compound fields and lower-hull three-phase
  facets pass.
- Al-7Si-0.3Mg Scheil: liquidus/range, eutectic fraction, and terminal Mg2Si
  event pass.
- Cu-Au ATAT-style CE: fitted ECIs recover L1₀ CuAu and both L1₂ ground states;
  all predicted ordering temperatures are within 50 K.
- Benchmark test files: thermodynamics 5 passed; cluster expansion 1 passed.

## Package verification

- Thermodynamics full Vitest suite: 33 files, 315 tests passed.
- Cluster-expansion full Vitest suite: 2 files, 23 tests passed.
- TypeScript library typechecks for both packages passed.
- ESLint on all changed production TypeScript passed with zero warnings.
- Scheil performance benchmark executes 2,000 steps over 10 components and 20
  candidate phases below the 5-second gate.

## Performance measurements

- Binary native equilibrium: 0.001668 ms/solve (target below 1 ms).
- 20×20×20 conventional-cell FCC canonical MC: 32,000 atoms verified; 672 full
  lattice sweeps/s measured on this host. The benchmark records the requested
  10,000 sweeps/s target without converting site attempts into misleading
  “sweeps.”
- The new batched common-tangent binding measures approximately 14–28× against
  the equivalent production TypeScript CALPHAD solver, satisfying the expected
  10–100× equilibrium range. Persistent-session canonical MC is measured against
  the exact TypeScript mirror with the composition retained explicitly in
  benchmark output, but the equiatomic result remains approximately 42× and
  therefore below the expected 50–200× range.

The MC performance items remain open: 672 full lattice sweeps/s does not satisfy
the explicit 10,000 sweeps/s target, and the stable equiatomic binding speedup
is below 50×. On this approximately 2.0–2.2 GHz core, the throughput target
requires 320 million proposals/s (only about 6–7 cycles each). A separate
proposal-only ceiling benchmark measures two random-site selections and an
equality check without any CE energy or Metropolis work, providing an executable
lower bound. In release mode it reaches approximately 98.4 million proposals/s,
only 31% of the required 320 million proposals/s before any thermodynamic work.
Consequently the full-sweep target is not locally attainable without changing
the definition to attempted moves, changing the single-core constraint, or
changing the target itself.
