fix(mps): guard variational tol cast against overflow to infinity

[ultimatile]

Summary

The variational (fit) MPO-MPS apply path cast its f64 tolerance into the scalar's real type with a bare NumCast::from(tol).expect(...) and no post-cast finiteness check. For f32-family scalars, NumCast::from maps an out-of-range but finite f64 (e.g. 1e300) to Some(inf) rather than None, so the convergence comparison |current - prev| <= tol * current becomes trivially true and the fit reports convergence after one sweep with an arbitrarily bad result. Two hardened copies of the same cast already existed (successive-randomized apply and an algorithms-crate helper); this consolidates all three into one shared primitive and closes the gap on the variational path.

Closes #472

Changes

• ariadnetor-core: new try_real_from_f64<T: Scalar>(f64) -> Option<T::Real> (NumCast::from then an is_finite() filter), exported at the crate root next to the existing scale-safe norm helpers.
• ariadnetor-algorithms: delete the crate-local copy and repoint every consumer (DMRG heff / heff_block_sparse validation / sweep, Krylov lanczos and lanczos_kernels, including a #[cfg(test)] helper) onto the core primitive.
• ariadnetor-mps successive-randomized apply: real_from_f64 now delegates to the core primitive, preserving its panic message.
• ariadnetor-mps variational apply: both the Dense and BlockSparse tolerance casts gain the finiteness check via the core primitive; the now-unused NumCast import is dropped.

Impact

• ariadnetor-core gains exactly one new public symbol, try_real_from_f64 (confirmed via cargo public-api).
• Variational apply with a tolerance that is finite in f64 but overflows the scalar's real type now panics at the cast instead of silently casting to infinity and mis-converging. Callers passing a representable tolerance are unaffected; the change only converts a previously-silent numerical fault into a fail-fast.

Test plan

• Moved the cast's unit tests into ariadnetor-core, covering f32, f64, and Complex against overflow, NaN, and infinity inputs.
• Added a Dense f32 #[should_panic] regression test on the variational path, mirroring the successive-randomized one, driving a 2-site f32 MPS with a 1e300 tolerance.
• cargo test across ariadnetor-core, ariadnetor-algorithms, and ariadnetor-mps passes; cargo make clippy (warnings denied, all targets) is clean.

Notes

The BlockSparse variational cast has no dedicated f32 regression test: both cast sites are the identical one-line construct and try_real_from_f64 monomorphizes independently of tensor storage, so the Dense test covers the shared finiteness-check-then-expect wiring, and the BlockSparse site's reachability is already exercised by the existing f64 BlockSparse variational tests.

@coderabbitai ignore

[Copilot]

Pull request overview

This PR hardens tolerance handling across MPO–MPS algorithms by centralizing the f64 -> T::Real cast behind a shared ariadnetor-core helper that rejects non-finite results (notably f64 values that overflow to ±inf when cast to f32). This prevents spurious “converged” reports in variational apply paths for f32-family scalars and removes duplicated cast logic across crates.

Changes:

• Add ariadnetor_core::try_real_from_f64::<T: Scalar>(f64) -> Option<T::Real> and re-export it from the core crate root.
• Migrate ariadnetor-algorithms (DMRG + Krylov/Lanczos) and ariadnetor-mps (successive-randomized + variational apply) to use the shared helper.
• Add a variational Dense f32 regression test ensuring unrepresentable tolerances panic instead of silently becoming inf.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
crates/ariadnetor-core/src/numeric.rs	Introduces the shared, finiteness-checked try_real_from_f64 conversion + unit tests.
crates/ariadnetor-core/src/lib.rs	Wires the new numeric module and re-exports try_real_from_f64 at the crate root.
crates/ariadnetor-algorithms/src/lib.rs	Removes the crate-local numeric module entry point.
crates/ariadnetor-algorithms/src/krylov/lanczos.rs	Switches tolerance casting to ariadnetor_core::try_real_from_f64.
crates/ariadnetor-algorithms/src/krylov/lanczos_kernels.rs	Uses the core helper for RNG-derived real/imag conversions and test helpers.
crates/ariadnetor-algorithms/src/dmrg/sweep.rs	Imports try_real_from_f64 from core instead of the removed local helper.
crates/ariadnetor-algorithms/src/dmrg/heff.rs	Uses the core helper for eigensolver tolerance representability validation.
crates/ariadnetor-algorithms/src/dmrg/heff_block_sparse/validation.rs	Same validation change for the block-sparse DMRG heff path.
crates/ariadnetor-mps/src/apply/successive_randomized.rs	Delegates the SRC cutoff cast to the core helper (preserving panic semantics).
crates/ariadnetor-mps/src/apply/variational.rs	Fixes variational tolerance casting (Dense + BlockSparse) to reject overflow-to-inf.
crates/ariadnetor-mps/tests/mps/apply_variational.rs	Adds a Dense f32 #[should_panic] regression test for unrepresentable tolerances.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated no new comments.