fix(hardening): EPS harmonization, mypy gate, Zhou+Prony guards, anchor tests

[marcohost33-maker]

Summary

Targeted hardening of LiouScope (4 candidates from a verified deep-dive) plus a debug pass. No anchor values changed. No destructive git ops (ref AGENTS.md Working-agreements §2 / 2026-05-16 data-loss incident — branch-create + commit + push only).

1. EPS harmonization (S) — diagnostics/nonnormality.py

The hard-coded 1.0e-300 divide-by-zero floor for the Petermann inner product |<l,r>|^2 is now the named, shared _consts.EPS_DIV. The value is kept: 1.0e-300 is a floating-point underflow floor (≈ smallest normal double), not a physics tolerance — collapsing it onto EPS_GAP=1e-10 would wrongly flag near-defective-but-finite modes as inf. Verified empirically that the canonical fixtures have min|<l,r>|^2 ≈ 0.91…0.99, far from any threshold ⇒ no anchor change.

2. mypy CI gate (M) — ci.yml

Removed continue-on-error: true. Fixed all 18 previously type-blind findings (mostly no-any-return from numpy ops → annotated return locals; one is_dataclass instance-narrowing in io/export.py). No behaviour change. mypy src/liouscope now exits 0.

3. Zhou guard (M) — _zhou.py

Adopted canonical EPS_DIV (division floor) + EPS_GAP (zero-eigenvalue check) instead of inline magic numbers, matching nonnormality. New tests:

• closed-form anchor (single-qubit pure dephasing, H=0, jump=Z, γ=0.5): dissipator spectrum {0,0,-1,-1} ⇒ gap=1 (hand-derived, numerically confirmed); diagonal ⇒ K=1; predictor collapses to t_lower = t_upper = log(1/eps)/gap = log(1000) = 6.9077….
• defective-mode guard: a near-defective mode (denom→0) must be skipped (not poison K_max→inf), so t_upper stays finite.

4. Prony fallback (S) — fitting/prony.py

np.linalg.lstsq / np.roots on a near-singular Hankel matrix could raise or yield garbage. Now wrapped in try/except (LinAlgError, ValueError) → emits RuntimeWarning + falls back to a safe _default_seed (strictly positive amplitude).

fails-before proof (stashed prony.py, ran new tests against pre-guard code):

FAILED test_prony_seed_non_finite_input_falls_back[nan]   (LinAlgError: SVD did not converge)
FAILED test_prony_seed_non_finite_input_falls_back[inf]
FAILED test_prony_seed_non_finite_input_falls_back[-inf]
FAILED test_prony_seed_all_zero_signal_has_positive_amplitude  (assert 0.0 > 0)
4 failed → all pass after the guard

Debug pass

Property-probed fitting/ + diagnostics/ with empty / single / NaN / inf inputs. Prony was the single clear, provable bug. Two observations left as-is (documented/defensible, not bugs): choose_model({}) → "M0" (documented baseline fallback); estimate_neff_geyer(all-NaN) → n (NaN propagation).

Anchor impact

pytest tests/test_anchors.py → 17 passed / 2 skipped (QuTiP), 0 anchor values changed.

Test plan

• ruff check src tests → exit 0
• mypy src/liouscope → exit 0 (18 → 0)
• pytest -q → 162 passed / 4 skipped (was 155; +7 new tests)
• pytest --cov=liouscope --cov-fail-under=80 → 87.95% coverage
• anchor regressions green, 0 anchors changed
• CI: 4 Python-matrix jobs (3.10–3.13) green — pending
• CHANGELOG.md updated (Unreleased section)

Local verification on Python 3.14 surrogate; CI matrix 3.10–3.13 is the real gate.

🤖 Generated with Claude Code