fix(mps): surface non-finite SRC apply state as ApplyError

[ultimatile]

Summary

The SRC apply path (ApplyMethod::SuccessiveRandomized) returned a compressed MPS as a success even when the computation had been poisoned by non-finite values — an overflowed contraction producing inf, or inf - inf / 0 * inf producing NaN downstream. This PR surfaces that state as an error at the result boundary: every growth round scans its sketch panel for non-finite elements, and every assembled site tensor is scanned before the state is returned, so apply_with_method returns Err(ApplyError::NonFinite) instead of a poisoned success, in both adaptive and fixed mode. Closes #481.

Changes

• crates/ariadnetor-mps/src/types.rs — new public ApplyError enum (thiserror, #[non_exhaustive]) with a NonFinite { site, norm } variant carrying lossy f64 diagnostics; site is documented as the detection locus, not the poison's origin.
• crates/ariadnetor-mps/src/apply/mod.rs — shared check_finite boundary predicate: an elementwise scan of each element's real and imaginary parts, rather than a norm fold, so a finite state whose Frobenius norm merely overflows the scalar's real type is not rejected.
• crates/ariadnetor-mps/src/apply/successive_randomized.rs — apply_successive_randomized_dense returns Result; per-round sketch-panel scans (fail-fast, both modes), a single assembled-sites scan before the state is built, and the same check on the n == 1 path.
• crates/ariadnetor-mps/src/dispatch.rs — the SRC kernel-dispatch method and apply_with_method return Result<_, ApplyError>; the other methods' arms wrap their result in Ok; the block-sparse SRC arm keeps its dense-only panic. apply stays infallible (the current default method has no failure path) with its doc contract made precise.

Impact

Compile-breaking for apply_with_method callers; all workspace consumers (the mps integration tests, crates/ariadnetor-mps/tests/authority.rs, and the algorithms mpo_mps_apply bench) are updated via apply_ok unwrap helpers. The other apply methods are behaviorally unchanged and still return a poisoned state as Ok, as documented on the error variant.

Test plan

• New error-path tests: NaN injected into an MPS site (adaptive mode), inf injected into an MPO site (fixed mode), and a NaN single-site product (n == 1 path) — each asserting the NonFinite diagnostics, including detection at the sweep's first processed site.
• Existing finite-input apply tests cover the Ok-path regression (now unwrapped).
• The assembled-sites scan has no deterministic trigger with finite panels: the gap between a finite sketch panel and a non-finite assembled tensor depends on the random Gaussian sketch, so a triggering fixture would encode magnitudes tied to the RNG stream. Recorded as a deliberate omission in the test file.
• cargo make gate (fmt-check, clippy -D warnings over all targets, workspace tests including doctests): clean; cargo test -p ariadnetor-mps: 264 passed.

Notes

• The scans run before the optional canonicalize + truncate finishing pass; non-finite values arising only inside that pass are the truncation machinery's concern.
• The adaptive stopping rule itself is not robust at numerical extremes — the running sketch norm can saturate to inf across rounds, and an overflowed R^-1 can collapse the error estimate to zero — and degrades toward premature convergence with a finite result; tracked in SRC adaptive stopping rule degrades toward false convergence at extreme scale #483.
• Per-method apply entry points with method-specific error types were considered and deferred: no non-test consumer selects an apply method statically today.

@coderabbitai ignore

[Copilot]

Pull request overview

This PR makes the successive randomized compression (SRC) MPO→MPS apply path fail fast when non-finite values (NaN/inf) reach defined result boundaries, returning a new ApplyError::NonFinite instead of silently producing a poisoned Mps. It updates apply_with_method to be fallible (Result) while keeping apply infallible since the default method remains non-failing.

Changes:

• Introduces ApplyError and threads Result<_, ApplyError> through the SRC dispatch and apply_with_method.
• Adds shared boundary finiteness checking (check_finite) and uses it within SRC per growth round and for assembled sites.
• Updates tests and benches to unwrap “finite-input” apply calls via helper wrappers, and adds SRC-specific error-path tests.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
crates/ariadnetor-mps/src/types.rs	Adds public ApplyError with NonFinite diagnostics and documents error scope/semantics.
crates/ariadnetor-mps/src/lib.rs	Re-exports ApplyError from the crate root.
crates/ariadnetor-mps/src/dispatch.rs	Makes apply_with_method return Result and wires SRC dispatch to return ApplyError.
crates/ariadnetor-mps/src/apply/mod.rs	Adds shared check_finite boundary predicate for dense tensors.
crates/ariadnetor-mps/src/apply/successive_randomized.rs	Converts SRC dense apply to Result and adds per-round + assembled-site non-finite scans.
crates/ariadnetor-mps/tests/mps/helpers.rs	Adds apply_ok helper to unwrap apply_with_method for finite-input tests.
crates/ariadnetor-mps/tests/mps/apply.rs	Switches tests to apply_ok where inputs are finite; keeps default-vs-explicit contract as value equality.
crates/ariadnetor-mps/tests/mps/apply_variational.rs	Switches finite-input tests to apply_ok.
crates/ariadnetor-mps/tests/mps/apply_variational_block_sparse.rs	Switches finite-input tests to apply_ok.
crates/ariadnetor-mps/tests/mps/apply_block_sparse.rs	Switches finite-input tests to apply_ok.
crates/ariadnetor-mps/tests/mps/apply_successive_randomized.rs	Adds SRC non-finite surfacing tests; keeps validation panic tests.
crates/ariadnetor-mps/tests/authority.rs	Adds local apply_ok helper to keep dispatch-count tests focused while unwrapping fallible apply.
crates/ariadnetor-algorithms/benches/mpo_mps_apply.rs	Adds apply_ok helper and updates benches to unwrap fallible apply on finite inputs.

---

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