CHANGELOG.md

Changelog

All notable changes to LiouScope are documented in this file. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Fixed

• D16 was published from the spectrum for which D1/D3/D4 had just been withheld (PR #121, round-23 review, finding 13). Where the zero-mode certificate is applicable but unresolved, the spectral layer reports D1, D3, D4 as NaN and has_complex_pairs as None, because the candidate spectrum is explicitly untrustworthy. diagnose() nevertheless passed those same eigenvalues into lep_proximity, and the LEP layer returned a finite D16. Measured on the issue-#113 stiff fixture: lep_proximity = 0.0 with 51 candidate pairs -- and 0.0 is not a neutral number but the coalescence limit, i.e. the strongest exceptional-point signal the diagnostic can emit, manufactured out of slow modes sitting below the eigensolver's backward error. A missing or ambiguous slow mode changes the closest eigenvalue pair directly, so the number was never a measurement.

  compute_lep_layer gains spectral_resolved (default True, so callers who say nothing keep measuring) and withholds D16 as NaN with lep_candidate_count=None when it is False; LepResult.lep_candidate_count is therefore int | None, in parity with SpectralResult.has_complex_pairs. diagnose() derives the flag from the same applicable and not resolved predicate the spectral layer uses, so the whole run withholds on one condition rather than on two. The withheld value is deliberately distinguishable from both measured extremes: 0.0 (coalesced) and inf (no pair) are answers, NaN is the absence of one, and _strip_unavailable in the classifier keys on exactly that. D17 and D18 stay measured -- D17 already inherits the withheld gap, D18 is computed from the operator.

  The certificate floor downstream caps the classifier's verdict, which is a different guarantee from not publishing the number: LepResult is returned to callers and persisted as audit metadata.

• A non-trace-preserving operator became certificate-applicable when it was written in small enough rate units (PR #121, round-23 review, finding 12). trace_preservation_defect squares its entries before summing, so for diag([0, -1e-200, -2e-200, -3e-200]) every square falls below the smallest subnormal and the pair comes back (0.0, 0.0). The applicability gate then evaluates 0.0 > tp_rtol * max(0.0, tiny), which is False for every operator, and both certificate APIs reported applicable=True, certified=True for an operator whose relative trace defect is 3 / sqrt(14) ~ 0.80. The same operator at 1e-150 was correctly refused, so a pure change of rate unit decided whether the object counted as a legal generator. Both norms are now recomputed on a copy scaled by the largest finite component whenever the reference scale has been lost.

  The rescue is deliberately one-directional: it repairs underflow and leaves overflow alone. ||L||_F = inf for the round-20 counterexample is not an accident to be worked around but the input to the round-21 refusal of a non-finite reference scale. That norm is mathematically about 1.4e308 and therefore representable, so a scaled computation would hand back a finite number, readmit the operator and silently reopen a hole that took five interpreter versions of CI to close. The two directions look symmetric and are not: only one of them has a decision behind it. A dedicated test pins the overflow refusal so that a later "symmetric" tidy-up goes red.

• The workflow hardening gate was blind to the commonest way to write a step (PR #129). .github/scripts/check_workflow_hardening.py enforces the SHA-pinning rule of AGENTS.md section 4 on every workflow, and nothing enforced it in turn. It reported green while four classes of unsafe workflow passed:

  • USES_RE was ^\s*uses:, which does not match the list form - uses:. Seven of 27 action references in this repository were invisible to the gate, every actions/checkout among them. They are correctly pinned today by discipline, not by this check; a PR moving one to @main passed.
  • docker:// was exempt outright, so a mutable third-party tag (docker://org/img:latest) was waved through. The exemption now requires an immutable @sha256: digest.
  • permissions was checked for presence only, so permissions: write-all -- a declaration of total access -- counted as evidence of least privilege.
  • The pull_request_target waiver was a substring test over the raw file, so writing the waiver's name inside a # comment satisfied it.

  The github.com/ exemption is removed as dead code: uses: does not accept that prefix and a GitHub owner name cannot contain a dot, so no such organisation can exist. It was the subject of a CodeQL incomplete-substring alert -- which is the only reason the file came under scrutiny. CodeQL flagged the harmless line and missed the regex two lines above it that actually broke the gate.

  Closing the docker:// hole initially rejected digest-pinned containers: _check_uses_pin splits the reference at @ before consulting _is_third_party_uses, so a digest test there ran on the truncated string and the digest then failed the 40-character git-SHA rule. Container digests and git SHAs are now checked as the different pin shapes they are. That regression was caught by an over-correction control, not by review.

• Two zizmor audits re-enabled whose suppression reasons no longer hold (PR #129). impostor-commit was disabled because the audit crashed on the private cross-repo pin coworkerz-ci; that pin is gone, and the only two occurrences of the name left in the tree are the comments explaining the suppression. It is the one audit that verifies a 40-character SHA actually exists upstream, so the entire pinning strategy was unverified without it. advanced-security: false was justified with "private repo without GHAS" -- this repository is public, measured rather than assumed, so zizmor findings were never uploaded to Code Scanning: absent there for want of an upload, not for want of findings. (annotations is set to false alongside it; the action refuses to start with both enabled.)

Added

• Tests for the hardening gate itself (PR #129). It had none. Each unsafe workflow is asserted next to a positive control that must still pass -- rejection alone would be satisfied by a gate that fails everything, which is exactly as useless as one that passes everything. Includes a digest-pinned container that must remain allowed, and a check that the real tree passes its own gate.

Known gaps

• quality contract is not a required status check on main; only test (3.10-3.14) and qutip-cross-check (3.11/3.12) are. Five security workflows run and none of them blocks a merge, so the comment in .github/workflows/zizmor.yml claiming the step "GATET die CI (rot = blockt merge)" is factually wrong. Making it required is the right follow-up, but only after this fix lands -- otherwise a blind gate becomes mandatory.
• Eight gates that decided from what they never had (PR #121 round-22 review). CHANGES StabilityReport PAYLOAD TYPES and the A10/F5 CLASSIFIER BRANCH. All eight share one shape: a value that is not a measurement is used as if it were one.
  • Non-finite tolerances were honoured instead of refused. defect > tp_rtol * fro is False for every operator when tp_rtol is NaN, and has an infinite right-hand side when it is inf, so certified_eigvals(diag([0,-1,-2,-3]), tp_rtol=nan) came back applicable=True, certified=True for a trace defect of 3.0. Round 21 closed this shape one variable further along (the reference scale fro) and left the parameter beside it unchecked; rtol and tp_rtol are now validated in certified_eigvals and certified_eig.
  • An inapplicable certificate reported one zero mode. zero_mode_count kept its dataclass default of 1 in both inapplicable branches, so diag([1,2,3,4]) -- a spectrum with no zero eigenvalue -- claimed a stationary mode into persisted audit metadata. It is now 0.
  • The refined tolerance underflowed to zero. sqrt(lo * hi) forms the product first, and that product scales as c^2 under a uniform rate rescale while the refinement itself is scale free; below c ~ 1e-139 the tolerance became 0.0 and the numerical stationary residual survived the gap filter as physics. sqrt(lo) * sqrt(hi) is the same number and cannot underflow (measured: lo=1e-177, hi=1e-165 gave 0.0, now 1e-171).
  • An overflowing spectral radius produced a gapless verdict. |1.3e308 + 1.3e308j| exceeds the double range, so np.abs returned inf although every component is finite and passed the finiteness gate. spectral_zero_tolerance returned inf, no mode satisfied |lambda| > tol, and D1/D3/D4 all reported 0.0 for a spectrum with modes of order 1e308. Now computed by scaling on the largest component (measured after: tolerance 4.08e295, D1 5.0e307, D3 1.3e308, D4 8.0e307); a tolerance that is still not finite is refused.
  • The sparse builder inherited none of the dense overflow repair. The round-18 fix divided each diagonal entry before summing in build_liouvillian only; build_sparse_liouvillian still called diagonal().sum(), overflowed to inf, and accepted the very matrix its dense twin rejected on the same input (Hermiticity defect 1e290). Both paths now share the overflow-safe gauge shift and the refusal of a non-finite derived scale.
  • The jackknife never asked whether its fits converged. The round-20 guard protects the bootstrap replicates only; _jackknife copied fit.params unconditionally, and a failed least_squares returns its unchanged starting value -- theta_hat itself. The BCa acceleration is a third moment of exactly these estimates, so the endpoints moved with no data behind them (measured: one non-fit in forty shifted one interval width to 0.923x). _jackknife now raises RuntimeError, which compute_relaxation_layer already routes to bca_ci_beta = (nan, nan).
  • The unresolved run was the one whose report would not serialise. build_stability_report copied D1/D3/D9 through as raw floats while the layers above deliberately publish NaN when a certificate is unresolved, and dump_stability_report writes with allow_nan=False -- so json.dumps raised on precisely the run whose audit record matters most. Non-finite diagnostic values are now encoded as null, the placeholder already used for D8b/D10b and by ZeroModeCertificate.as_dict. Consumers reading diagnostics["D1_gap"] must accept None.
  • A missing gap was read as the gapless limit. _f5_reach defaulted gap to 0.0, which is positive evidence for the phantom-relaxation reach leg. With an unresolved certificate the spectral layer withholds D1 as NaN, _strip_unavailable removes the key, and with henrici_eta > 1 the classifier returned A10/F5 and marked the hypothesis SUPPORTED for a reach it could not compute. The downstream certificate floor caps the verdict at UNDEFINED but does not withdraw the class, the family or the matrix status. gap is now a REQUIRED key of that condition, so the rung is UNEVALUABLE without it; a gap MEASURED as 0.0 still fires the documented gapless branch. This reverses the deliberate choice recorded in test_missing_optional_key_is_reported_without_forcing_unevaluable, whose invariant now travels on gap_to_gns_ratio -- a key that genuinely has documented semantics without a value.
• The second repair step ended the eigenvalue ladder instead of continuing it (PR #121 round-18 review). certified_eigvals guarded its primary solve but not the dgeev-real step that runs next -- and unlike the primary, that step sits BEFORE zgees-schur and balanced-zgeev. A LinAlgError there left the generator and discarded two repair routes that could still have certified the spectrum. The sibling ladder in certified_eig already suppressed exactly this step, so the defect was two ladders drifting apart, not an overlooked case. A structural test now fails if either dgeev-real route is left outside a suppression block.
• The anti-overfit gate scored fits that never converged (PR #121 round-19 review). CHANGES HoldoutResult.accept for saturated fits. holdout_validate ignored fit.success and scored fit.params regardless. On the saturation plateau the model is CONSTANT, so train and holdout RMSE are equally enormous and their ratio is about 1 -- measured 1.0106 on holdout_validate(M0, linspace(0, 1e10, 64), exp(-5t/1e10), [1, -1]), which passed the 1 + delta criterion and returned accept=True. The gate that exists to catch non-generalising models was satisfied by a non-model. accept now requires fit.success, and the new HoldoutResult.fit_success field keeps "did not generalise" separable from "was never fitted" -- they demand different responses.
• D3, D4 and the oscillating-pair flag are withheld with D1 (PR #121 round-19 review). CHANGES SpectralResult OUTPUTS when an applicable certificate stays uncertified. The round-17 repair withheld D1 and stopped there; oscillating_gap, spectral_spread and has_complex_pairs were still published from the same candidate spectrum the layer's own warning calls untrustworthy. Withholding D1 alone was not a partial fix but an inconsistent one: it taught consumers that this layer withholds what it cannot stand behind, which made the surviving finite values more credible, not less. has_complex_pairs is now bool | None, because a bool has no NaN and False would assert "no oscillating modes" rather than report an unanswered question; classification maps None to the evidence dict's NaN sentinel, where _strip_unavailable removes it and the A8 rung does not hold.
• A comparison against NaN read as "no Hermiticity violation" (PR #121 round-19 review). np.trace(H) sums the diagonal before dividing, so it overflows to inf for a FINITE H with large entries -- the explicit finiteness gate immediately above has already passed at that point. The gauge-fixed scale then became NaN and defect > EPS_HERMITICITY * scale evaluated False, so build_liouvillian accepted a two-dimensional 1e308 * I carrying a one-sided 1e290 entry, an order-one gauge-fixed Hermiticity defect. The gauge shift is now computed as sum(diag(H) / d), which is bounded by max|diag(H)| and cannot overflow while H is finite; independently, a non-finite derived scale is refused outright so that any future route to one cannot silently reopen the hole. A genuinely Hermitian 1e308 * I is still accepted -- the repair is not a size limit.
• One zero-mode cutoff, obtained one way (PR #121 round-17 review, four findings). CHANGES NUMERICAL RESULTS on any generator where the a posteriori refinement rescues a genuine slow mode. Four consumer layers kept filtering with the certificate's raw band after the refinement had lowered the applied tolerance, discarding the rescued mode one layer after the certificate saved it. Measured on the two-level omega = 1, rates 1e-15 / 1e-14 generator (bound = 2.22e-13, applied = 5e-16, ratio 444): D24 gap 2.05e-14 -> 1e-15 and the mixing-time window 3.37e14 -> 6.91e15; D9 kept 2 modes with petermann_max = 1.0 and now keeps 3 with 2.0; the D19 overlap moved from 0.0 — the false A11/F4 trigger on the highest rung — to 0.141, and expansion_alpha from -1e-15 to -33.6; the D11 fallback scanned 12 of 15 certified modes and now scans all 15. These were not four independent mistakes: the same two-branch expression was copy-written at five sites, only the spectral layer was migrated when the refinement landed, and the docstring of operator_zero_tolerance still told readers to filter "equivalently, with the bound of the ZeroModeCertificate" — true when written, false since. The decision now lives once, in ZeroModeCertificate.zero_set_tolerance(); bound is documented as report-only; _certified_decomposition hands its consumers the certificate instead of a bare number; and a guard test pins the remaining .bound readers so a sixth site fails the suite instead of a review.
• D1 is withheld when no repair route certifies the spectrum (PR #121). With an applicable certificate and certified=False, compute_spectral_layer still published a finite gap read off the explicitly untrustworthy candidate spectrum. The warning does not reach a caller that consumes SpectralResult.gap, and _gather_evidence derives ratios from that value, so the number travelled further than its caveat. gap is now the NaN unavailable sentinel for both unresolved kinds, as the ambiguous case already was. D3/D4 are unchanged (see the note in spectral.py).
• A saturated GLS fit is no longer selectable (PR #121). Flipping GLSFitOutput.success changed nothing: no consumer read the flag, the fit still carried a finite AICc, and the exact plateau case the guard detects could win choose_model and supply the reported decay rate. The failure is made non-selectable at the single choke point (_fit_with_model assigns aicc = inf) rather than as four separate obligations on the consumers; with no selectable model left, aicc_model is the existing "none" sentinel and beta_D is NaN instead of being read off a failed fit. parametric_bootstrap refuses a saturated base fit outright — a resample around a non-estimate is not an uncertainty — and reports the count of non-converged replicates it retained (retaining widens the interval; dropping them would narrow it).
• The eigensolver repair ladder survives a primary nonconvergence (PR #121). A LinAlgError from the incumbent zgeev call ended certified_eigvals / certified_eig before the real-driver, Schur and balanced routes were tried — defeating the ladder in precisely the case it exists for. The error is now carried and re-raised only if no route produced a spectrum.
• The zero_mode_certificate field documentation matched neither the code nor this changelog (PR #121). It described a report-only field while classify_mechanism reads it and _apply_spectral_certificate_floor caps both the reported verdict and the tier from it. Documentation only; no behaviour change.
• The zero-mode band gets a second, independent axis: an a posteriori backward-error certificate (issue #118 finding 15). CHANGES NUMERICAL RESULTS on generators whose slowest decay sits far below the scale their Hamiltonian sets. The band is 1e3 * eps * ||L||_2, and ||L||_2 is fixed by the oscillation frequency while the decision is about decay: a two-level system with omega = 1 and jump rates 1e-15 / 1e-14 had its genuine slowest mode (D1 = 1e-15) swallowed by the band, was certified resolved with zero_mode_count = 2, and D1 reported the next, twentyfold faster eigenvalue (2.05e-14) with full confidence. The magnitude axis cannot fix this, and the ZERO_MODE_AMBIGUITY_FACTOR comment already said why: healthy round-off reaches 2.38 * eps*||L|| while unresolved slow modes were measured at 4.87 — populations a factor of two apart, which no threshold separates. certified_nonzero_modes therefore adds the per-mode bound |lambda - lambda_hat| <~ max(||L x - lambda_hat x||, ||L^H y - conj(lambda_hat) y||) / |y^H x|, which is a certificate: were the mode stationary, the bound would force |lambda_hat| <= bound, so exceeding it proves it is not. Measured over the same corpus that calibrated the existing factor (96 healthy generators / 132 in-band modes; 6 stiff generators with analytic gaps): 0 healthy modes reached q = 1 (max 0.472) and all 3 stiff members carrying a genuine in-band mode were rescued. The refinement is one-directional — it can only take a mode OUT of the zero set — so it cannot produce the false "unresolved" verdict the ambiguity split guards against. The certificate now carries zero_tolerance, which D1/D3/D4 filter by; without a rescue it equals bound and the healthy path is unchanged bit for bit.
• A GLS fit that ends inside the model magnitude guards fails closed (issue #118 finding 9). The guards keep an out-of-range optimiser probe finite, which is what they are for — but the value they return is constant, so its derivatives vanish and least_squares terminates on "gradient is small". Measured: fit_gls_ar1(M0, t in [0, 1e10], y = exp(-5t/1e10), p0 = [1, -1]) returned success=True with p0 unchanged and a residual norm of 7.9e100. Only the FINAL evaluation is judged (probes passing through the plateau are exactly what the caps exist for); GLSFitOutput gains saturated, naming which guard fired.

Documentation

• SpectralResult.zero_mode_certificate is no longer described as "report-only" (issue #118 finding 16). It was load-bearing from the commit that introduced _apply_spectral_certificate_floor, and is now load-bearing on a second path through zero_tolerance. See the corrected entry below.

Fixed

• Prony fallback seeds are grid-relative (round-16 review; the #108 class of defect). CHANGES NUMERICAL RESULTS on grids far from unit span where the Prony estimate falls back (non-uniform sampling, short signals, degenerate data). The fallback seeded (beta, omega) = (1, 1) in absolute rate units, so on a valid non-uniform grid spanning t = 1e7 the M3b fit started seven orders of magnitude off and least-squares "converged" (success=True) onto the seed itself — measured [A, beta, omega] ≈ [0.006, 1, 1] for true values 5e-8 / 2e-6 — corrupting the fitted rate and suppressing the M3b/A8 hypothesis purely because of the rate unit. Fallback rates are now 5.0 / t_span and the success-path positivity floors 5e-6 / t_span, both reproducing the historical values exactly at the canonical t_span = 5 (the same convention as ALPHA_SEED_FLOOR_FRAC / M3A_SLOPE_SEED_FRAC); the M3b fit on the measured non-uniform scenario is now unit-invariant across c ∈ [1e-6, 1e6].
• D24 uses the eigenvalue certificate when only the gap is missing (round-16 review). The recomputation branch required certified_eig even when the caller supplied petermann_factor — but that partial path consumes no eigenvectors, so a stiff network whose eigenvalue certificate resolves a usable gap (measured 3.32e-6) while only the round-15 eigenvector gate fails returned an unnecessary unconverged record. The certificate now matches what is consumed: certified_eigvals when only the gap is recomputed, the stricter certified_eig when the Petermann factor is.
• D11 is decoupled from the D9 vector gate (round-16 review). When D9 is withheld (round-15), the NaN sentinel array flowed into bohr_arithmetic_progression, which silently filtered it and reported the default length 1 as a measured D11 value. D11 consumes only eigenvalues, and the eigenvalue certificate frequently still resolves in exactly that situation, so its input is now recomputed from the certified spectrum; only when the eigenvalues themselves are unresolved does bohr_ap_length become NaN (the field is now float-typed with NaN as the documented unavailable sentinel).
• Release-note correction (round-16 review, documentation only): the atol= opt-in for the legacy absolute Hermiticity gate exists on the standalone is_hermitian predicate only; the builders never accepted it and their relative gate cannot be bypassed. The #109 entry below now says so explicitly instead of advertising an unavailable builder argument.
• certified_eig validates eigenVECTOR residuals before certifying (round-15 review). CHANGES NUMERICAL RESULTS on stiff generators. The certificate accepted a decomposition solely because one eigenvalue was close to zero — but D19 and the Petermann factors consume the eigenvectors, and a small |lambda| does not vouch for them. Measured on a valid stiff classical network: certified=True, resolved=True while the slow mode's LEFT eigenvector had residual three orders of magnitude beyond the certificate bound, i.e. not an eigenvector in any usable sense (the right vectors were fine). Acceptance is now per mode and relative — r_j <= max(VECTOR_RESIDUAL_REL_MAX * |lambda_j|, bound) with the larger of the unit-normalised left/right residuals — because r/|lambda| is the first-order relative error scale of anything computed from the pair, and a single operator-scale cutoff does not separate the measured populations (healthy <= 2.1e-10, legitimate dgeev-real repairs a few percent, corrupt decompositions 22%–2900%; the calibration is documented at the constant). A candidate failing the gate does not end the ladder; when no route passes, the certificate reports certified=False with the offending vector residual, and every vector consumer (D19, D9, D24) withholds. certified_eigvals is deliberately untouched by this gate: the eigenvalues of such a decomposition remain usable for D1/D3/D4.
• The repair ladder continues past ambiguous candidates (round-15 review). CHANGES NUMERICAL RESULTS on stiff generators, in the fail-open-to-correct direction. A candidate whose in-band spectrum contained an ambiguous mode (#113) previously ended the ladder immediately with resolved=False — withholding D1 and flooring the verdict — even when the very next route resolves the generator cleanly (measured: zgeev certified with one ambiguous mode at 6.3e-7 while dgeev-real returns a 2.1e-15 stationary mode with none). Both ladders now accept only ambiguity-free candidates and fall back to the best ambiguous candidate (fewest ambiguous modes) only when every route stays ambiguous. The healthy path is unchanged and still lazy.
• Petermann factors (D9) consume the certified eigendecomposition (round-14 review). CHANGES NUMERICAL RESULTS on stiff generators. petermann_factors recomputed its own raw zgeev decomposition, so on a stiff trace-preserving generator it consumed exactly the solver failure the spectral and Mpemba layers repair (#112) — and no zero-mode cutoff can restore an eigenvalue that is absent from the raw spectrum. Measured on the stiff four-level fixture: 16 "non-zero" modes with petermann_max ≈ 622.7 against the certified 15 modes with ≈ 2.0, corrupting D9 and the D11 input while the spectral certificate looked resolved — a possible false F1 signal no verdict floor would catch. The function now uses certified_eig; when the certificate is applicable but unresolved (#112/#113) both returned arrays are the NaN unavailable sentinel, so petermann_max becomes NaN and the F1 rung reads UNEVALUABLE instead of the empty-array default K_max = 1.0, which would assert perfect normality.
• Radius fallback when the zero-mode certificate is inapplicable (round-14 review). Without established trace preservation no zero eigenvalue is guaranteed, so the operator-norm certificate bound is not a valid zero-mode cutoff: on a non-trace-preserving, strongly non-normal input the bound (measured 2.2e3) exceeded the whole spectrum {-1,-2,-3,-4} and the round-13 filters discarded every eigenvalue — D1 = 0.0 (gapless, firing the F5 reach leg) and an unconverged D24 for a well-separated spectrum with true gap 1. All round-13 call sites (spectral layer, Mpemba layer, D9, D24) now use the certificate bound only when certificate.applicable, and fall back to the radius-based #108 tolerance otherwise.
• One zero-mode scale for certification and filtering (round-13 review of the #112 machinery). CHANGES NUMERICAL RESULTS on strongly non-normal generators. The ZeroModeCertificate accepts a stationary residual up to the true eigensolver backward error rtol * eps * ||L||_2, but the downstream zero-mode filters (spectral_zero_tolerance, #108) only see the spectrum and used the spectral radius max|lambda| as a proxy. For a strongly non-normal trace-preserving operator ||L||_2 exceeds the radius by orders of magnitude (measured: 3.9e3 on a 4x4 example), so a certified-resolved zero mode with residual between the two thresholds survived the radius filter as a spurious "genuine" mode: D1 reported ~1e-12 — occasionally negative, impossible for a GKSL generator — instead of the true gap 1, and D9 reported four eigenmodes where three exist. Every consumer that holds the operator now filters with the certificate's own bound (new shared helper liouscope.numerics.linalg.operator_zero_tolerance): the spectral layer (D1/D3/D4, has_complex_pairs), the Mpemba layer (slowest mode, mode expansion), petermann_factors (D9) and the D24 predictor. Genuine slow modes inside the coarser band are not silently swallowed — the #113 ambiguity split reports them as resolved=False and floors the verdict. The radius-based default of spectral_zero_tolerance is unchanged for spectrum-only call sites, and a caller-supplied atol still wins.
• D24 recomputation is routed through the certified eigensolve (round-13 review). CHANGES NUMERICAL RESULTS on stiff generators. When compute_zhou_predictor recomputes an omitted gap/petermann_factor it filtered the raw zgeev spectrum, retaining exactly the solver failure the spectral and Mpemba layers repair (#112): on the stiff four-level fixture D24 reported gap = 7.28e-6 where the certified solve recovers the physical 1.074e-5, shifting the whole predicted mixing-time window by ~30%. The recomputation now uses certified_eig and the certificate's zero-mode bound; when the certificate is applicable but unresolved (#112/#113) the predictor returns an honest unconverged record (infinite bounds, NaN gap/K) instead of bounds built on a demonstrably wrong spectrum. Caller-supplied values keep bypassing the eigensolve entirely.
• Hypothesis matrix: a conclusively refuted partial rung is NOT_SUPPORTED, not UNEVALUABLE (round-13 review; #102 machinery). The rungs are conjunctions, so one evaluated-false condition refutes a rung no matter what a missing sibling measurement would have said (kreiss = 1 refutes F1 with or without petermann_max). Previously any missing required key forced UNEVALUABLE, which also propagated into an UNEVALUABLE A12 fallback while the decision ladder deterministically returned A12 — the matrix contradicting the decision it documents. UNEVALUABLE is now reserved for the genuinely open case: no evaluated condition is false and the missing evidence could still flip the rung to supported. Absent keys stay listed in missing for the audit trail either way; claim floors follow the status as before (refuted → NOT_EXCLUDED, open → UNDEFINED).
• D1 refuses to report a gap it cannot resolve, instead of reporting a fast mode (issue #113, PARTIAL). CHANGES NUMERICAL RESULTS on very stiff generators, where the previous value was wrong by orders of magnitude. The #108 tolerance carries a safety factor of ZERO_MODE_EPS_FACTOR = 1e3 above the bare backward error eps * ||L||. Once the spectral spread grows large enough, genuine slow modes fall inside that safety factor; the filter discards them as "zero" and D1 reads off the next surviving mode — a fast one. Measured on a four-level classical jump network: 5.0e7 reported for a true gap of 1.07e-5, i.e. wrong by twelve orders of magnitude and in the fail-open direction. ZeroModeCertificate now separates the two readings of an in-band mode: a degenerate stationary manifold (conserved quantity / symmetry sector) puts its extra zero modes at machine zero and is perfectly legal, whereas a swallowed slow mode sits above eps * ||L||, inside the safety factor. Only the latter (ambiguous_count > 0) marks the layer unresolved, and D1 is then reported as NaN — deliberately not 0.0, which asserts gaplessness and fires the F5 reach leg, and deliberately not the fast mode. ZERO_MODE_AMBIGUITY_FACTOR = 30 is calibrated against a measured distribution rather than chosen: across 83 healthy generators the largest genuine in-band |lambda| / (eps*||L||) reaches 2.38 (median 0.39), while unresolved slow modes were measured at 4.87 and 4.87e2. The two populations overlap within about 2x, so the split sits an order of magnitude above the healthy maximum rather than between them: a false NaN destroys a correct analysis, a missed marginal case only leaves the previous behaviour. Bounded reach, stated explicitly: roughly one further decade of spread is now fail-closed; beyond that the slow modes sink below the backward error and the defect is undetectable by any magnitude test. Issue #113 stays open for the extended-precision / Krylov route.
• Spectral layer: the computed spectrum is now checked against an exact structural fact before D1/D3/D4 are read off it (issue #112). CHANGES NUMERICAL RESULTS on stiff generators. Every trace-preserving generator satisfies vec(I)^H L = 0 exactly, so 0 is an exact eigenvalue and a correct eigensolve must return some |lambda| <= ZERO_MODE_EPS_FACTOR * eps * ||L||. A spectrum without one is proof that the solve failed — a theorem, not a tuned threshold. This is a different defect from #108: #108 fixed the zero-mode filter, while here the spectrum being filtered is wrong, so no separation tolerance can repair it (asserted in tests/test_spectral_certificate.py: the relative and the legacy absolute filter return byte-identical wrong gaps). Mechanism: LAPACK zgeev deflates a subdiagonal when the Ahues-Tisseur test |h[i,i-1] h[i-1,i]| <= eps |h[i,i]| |h[i-1,i-1] - h[i,i]| passes. On a stiff generator the large diagonal entries inflate that bound, so the entire slow block deflates to its own diagonal. Measured on a four-level classical jump network (rate spread ~1e10, exactly trace preserving): the exact zero mode vanished from the returned spectrum, a spurious mode appeared in its place, and D1 reported 7.28e-6 for a generator whose true gap is 1.074e-5 (independently confirmed against the analytically decoupled population and coherence blocks, and against a matrix-exponential decay oracle). Note this failure is invisible to conditioning: the per-eigenvalue condition numbers of the wrong spectrum are 1–25, i.e. the solver reports the wrong answer as well conditioned, and explicit balancing does not fix it (both measured). It is also not repairable by deflating vec(I) with a dense similarity: that destroys the block structure which made the slow modes computable in the first place, and was measured to make 12% of stiff systems worse. What works is re-solving the same matrix by a route whose deflation is not driven by the stiff diagonal. numerics.linalg.certified_eigvals() therefore tries zgeev, then (when L is real-valued) dgeev, then the complex Schur form, then a balanced zgeev, and returns the first spectrum satisfying the certificate. The incumbent result is kept byte-identically whenever it is already correct — measured across 400 random four-level networks, stiff and well-conditioned alike, zgeev was retained in 400/400 and no system's gap got worse. If no route is certified, compute_spectral_layer emits a RuntimeWarning and marks the layer unresolved rather than reporting a gap it cannot stand behind. SpectralResult.zero_mode_certificate records the outcome. The field is additive (it has a default, so the run-manifest contract is unchanged), but it is not report-only: classify reads certificate["resolved"] and applies _apply_spectral_certificate_floor, which caps verdict and tier. An unresolved certificate therefore changes the reported classification, by design — a certificate that could not withhold a verdict would be decoration. This entry previously called the field "report-only"; that description was wrong from the commit that introduced the floor, and the correction is recorded here rather than silently dropped. Since the a posteriori refinement below, the certificate also carries zero_tolerance, which D1/D3/D4 filter by — a second load-bearing path.
• Hermiticity validation is scale-relative: H is no longer accepted or rejected on the basis of its units (issue #109). is_hermitian and both Liouvillian builders applied an absolute 1e-9 tolerance to the Hamiltonian, which carries energy/rate dimension — the same defect class as #108, in input validation rather than in a diagnostic. Measured in both directions:
  • fail-open (the serious direction) — a genuinely non-Hermitian H with a relative defect of 1e-6 passed validation once ||H|| <~ 1e-3, producing a generator that is not GKSL at all, after which every downstream diagnostic describes dynamics that are not a quantum channel. The sparse builder shared the defect, so both paths accepted it.
  • false rejection — an exactly Hermitian H reconstructed by a unitary similarity carries round-off ~eps*||H||, which exceeds 1e-9 once ||H|| >~ 1e8, so valid input was refused. EPS_HERMITICITY is now read as a relative tolerance on max|H|, so at unit scale — the regime of every existing fixture — the gate is unchanged; only operators far from unit scale move. The legacy absolute gate remains available as an explicit atol= opt-in on the standalone is_hermitian predicate only (mirroring how #99 and #108 preserved their predecessors); the builders build_liouvillian / build_sparse_liouvillian deliberately expose no such override — their gate is always relative, and pre-validating with the predicate does not bypass it. is_density_matrix keeps its absolute reading: a density matrix is trace-normalised, so that reading is already relative to a fixed scale and the rate-dimension argument does not apply.
• Zero-mode separation is scale-relative: D1/D3/D4/D9/D16/D19/D20/D24 no longer depend on the choice of rate unit, removing the ZERO-MODE-INDUCED verdict flips (issue #108). CHANGES NUMERICAL RESULTS for generators whose spectral radius is far from unity. This does not make the mechanism verdict unit-invariant in general: the A10/F5 branch still gates on the rate-dimensioned henrici_eta > 1.0, so rescaling can still cross that threshold. That remaining limitation is tracked in #101 (slice C) and stated in the README; #108 removes a different, independent source of unit dependence. The filter deciding which modes are the steady state used an absolute floor |lambda| > 1e-10, duplicated across five modules. A Liouvillian carries rate dimension, so this broke in both directions under L -> cL:
  • small c — every genuine mode fell below the floor, so D1 collapsed to 0.0 (unconditionally firing the gapless F5 reach leg) and the D19 slowest-mode overlap collapsed to 0.0, raising a false A11/F4 Mpemba candidate on the highest-priority rung of the classifier. Measured end-to-end on an amplitude-damped qubit with rho_0 = |+><+|: the verdict moved from A12/none NOT_EXCLUDED EXPLORATION at c = 1 to A11/F4 CANDIDATE CONFIRMATION at c = 1e-10 — a textbook system promoted to a quantum-Mpemba candidate by a change of units alone. The issue-#68 non-triviality guard could not correct it, since it depends on the same _slowest_mode helper.
  • large c — the round-off zero mode (~eps * ||L||) rose above the floor and was counted as genuine, yielding a negative spectral gap (-1.4e-6 at c = 1e10 for a Rabi-driven damped qubit), which is impossible for a GKSL generator by definition. The correction is canonical and carries no free parameter: the new numerics.scale.spectral_zero_tolerance() derives the threshold from the spectrum itself as ZERO_MODE_EPS_FACTOR * eps * max|lambda| (spectral radius — homogeneous of degree one and unitary-similarity invariant, like rate_scale), with eps fixed at DOUBLE-precision machine epsilon: the library's own dense solves are guaranteed double (below), so the storage dtype of a spectrum says nothing about the precision it was computed in. A spectrum from a genuinely single-precision external/GPU solver is the one case needing a coarser threshold, is indistinguishable from a downcast double result by inspection, and therefore belongs to the caller: the eigenvalue-based diagnostics (D1/D3/D4/D16) expose the rtol multiplier for exactly that, so such callers can widen the filter without reverting to an absolute floor. All 21 anchor regressions and the full suite stay green, but the change is not confined to rescaled generators: at unit scale the threshold is ~2.2e-13 rather than the historical 1e-10, so a mode between those two values is now classified as genuine where it was previously discarded. That direction is the intended improvement — it is what rescues metastable slow modes — and no anchor system carries a mode in that band, which is why the reference behaviour is unchanged. Non-finite spectra now fail closed rather than silently reporting "no non-zero modes" — including on the legacy path, since a compatibility switch may restore the old threshold but must not restore silent acceptance of corrupted solver output. The legacy absolute floor remains available as an explicit atol= opt-in on every affected function, mirroring how #99 preserved the pre-#97 steady-state tolerance. The double-precision solver contract is now enforced, not assumed: eig_nonhermitian documents "always LAPACK zgeev", but scipy.linalg.eig dispatches by dtype and genuinely ran single-precision cgeev on a complex64 input (measured: ~30x the eigenvalue error of the double solve on the same stored matrix; numpy.linalg by contrast always computes in double and only casts the result back). Since every backward-error tolerance above is calibrated against the double solve, all dense eigen/Schur boundaries (eig_nonhermitian, the Mpemba layer, Petermann/Henrici, D24) now promote to complex128 before solving. Representation error already present in caller-supplied single-precision data is the caller's data quality and is neither masked nor "corrected". This supersedes the storage-dtype-derived epsilon briefly present during review (it discarded resolved metastable modes: a complex64 two-channel generator with rates 1.0/1e-4, true gap 5e-5, reported 5e-1). D1 deliberately does not clamp the gap at zero: after the scale-relative filter a positive Re(lambda) is no longer round-off but a genuinely unstable (non-GKSL) mode, and masking it would trade one silent failure for another. The threshold is a multiple of the eigensolver BACKWARD ERROR (ZERO_MODE_EPS_FACTOR * eps * max|lambda|), not a fixed fraction of the spectral radius. A fixed fraction would impose a dynamic-range ceiling and discard the genuine slow modes of a METASTABLE generator (A5): with 1e-10 * max|lambda|, two damping channels at rates 1.0 and 1e-12 (true gap 5e-13) reported a gap of 5e-1. Calibration measured across amplitude damping, Rabi-driven damping, dephasing, a strongly non-normal near-defective generator and a 64-dim 3-qubit chain, each at c in {1, 1e6, 1e12}: the numerical zero mode never exceeded 1.94 * eps * max|lambda|, so the default factor keeps ~500x headroom while resolving genuine modes further down than the pre-#108 absolute floor did at unit scale. Pinned by tests/test_zero_mode_scale.py (tests over c in {1e-10 ... 1e12}: dimensionless quantities invariant, rate-valued ones scaling by c, genuine metastable slow modes preserved, no false Mpemba candidate, no negative gap, and a discrimination test proving the atol opt-in really restores the old defect). The suite also pins the known non-invariance of the mechanism class (A10/F5 still gates on rate-dimensioned henrici_eta, issue #101), so the limitation stays visible instead of being mistaken for a passing invariance claim.
• Fit seed rate is grid-relative, not an absolute floor (issue #108 class, fitting path). CHANGES NUMERICAL RESULTS for fits on time grids far from unit span. initial_guess_m0 floored the seeded decay rate at an absolute 1e-3. A rate is 1/time, so on a grid spanning t = 5e10 — exactly what a slow generator in small rate units produces — the seed sat 1e7 above the true rate, in a region where exp(-alpha t) has underflowed flat and the optimiser has no gradient. The fit then returned the floor itself as the measured rate: beta_D == beta_D_linear == 1e-3 on a system whose true rate was 5e-11, silently corrupting D5/D17 while every spectral quantity looked healthy. The floor is now a dimensionless decay depth over the fitted window (ALPHA_SEED_FLOOR_FRAC / t_span, the rate that decays ~0.5 % across the grid), which reproduces 1e-3 exactly at the t_span = 5 used throughout the suite — so seeds there, and the anchors, are unchanged. The too-few-positive-samples fallback rate is likewise grid-relative (~one e-folding across the window) instead of an absolute 1.0. Model evaluations are additionally bounded in magnitude: capping the exponent bounds exp, but M3a multiplies it by (A + B t), whose magnitude is unbounded in the parameters, and least-squares then squares the product inside its own normal equations — overflowing in SciPy's trf rather than in this module. Residual limitation, deliberately not claimed as fixed: at c = 1e-10 the least-squares convergence criteria stop tracking the rescaling, so fitted rates are asserted invariant only over c in [1e-6, 1e10]. Tracked in its own issue; non-dimensionalising the fit is the proper resolution.

Changed

• Import convention in AGENTS.md corrected to the codebase (issue #110, documentation only, no code change). The rule demanded absolute intra-package imports; the package has never followed it — measured at 135 relative import statements across 39 of 53 modules and 0 absolute ones. Relative intra-package imports are standard for a src/ layout, so the convention was corrected rather than ~30 modules rewritten. The divergence's only effect was that automated reviewers kept flagging conforming code as a violation (which is how it surfaced, during the PR #107 review).

Added

• Hypothesis-wise evidence matrix + support_score + reachability gate (issue #102, additive, claim_status: pending, no verdict change).

  • ClassificationResult.hypothesis_matrix reports one entry for every hypothesis of the A1-A12 taxonomy: each decision rung, the A12 fallback, and the schema-reserved A6/A7/A9. Each entry records status (SUPPORTED/NOT_SUPPORTED/UNEVALUABLE/RESERVED), the supporting conditions with the evidence values they read, the failing conditions as counterevidence, missing required evidence keys, an explicit fail-closed claim_floor and the per-class ordinal support_score. Claim-floor rules: RESERVED/UNEVALUABLE → UNDEFINED (no rule / no evidence — no claim); NOT_SUPPORTED → NOT_EXCLUDED (a threshold that did not fire is absence of support, not proof of absence, issue #70 A5); SUPPORTED → the verdict the hypothesis would receive were it the winner (via the unchanged _confidence → _pick_verdict_tier → A11 maximally-mixed-floor pipeline), so the winner's floor equals the reported verdict exactly (pinned by test).
  • To keep the matrix from drifting away from the decision, the priority chain is now defined declaratively: _ladder_spec() returns rungs of atomic _Condition predicates (each naming the evidence keys it reads), and _hypothesis_ladder (winner + shadow report) and hypothesis_evidence_matrix are both derived from that one spec. Behaviour-preserving: conditions, order, short-circuit evaluation and all (class, family, verdict, tier, confidence) outputs are unchanged (669-test baseline green, anchors untouched).
  • ClassificationResult.support_score (issue #102 "rename" option 1): the honestly-named twin of confidence, with documented, explicitly NON-probabilistic ordinal semantics (0.20 < 0.50 < 0.70 < 0.85 < 0.95 ranks rule strength; no calibration evidence exists). confidence stays as the legacy alias carrying the identical value (pinned equal by test); a calibrated replacement remains gated on the preregistered validation design in issue #102. Additive + defaulted (NaN) so serialised older results stay valid.
  • Reachability/ontology gate: new liouscope.REACHABLE_A_CLASSES (taxonomy minus RESERVED_A_CLASSES, 9 classes) is the coverage denominator for any "n of N classes" statement; reserved classes appear in the matrix as RESERVED with a permanent UNDEFINED floor and are excluded from claims. RESERVED_A_CLASSES is now exported too.
  • The per-hypothesis numerical-uncertainty and perturbation-robustness columns from the issue-#102 wishlist are deliberately not faked; they remain open in the issue. No MANIFEST_SCHEMA bump: the run-manifest contract is untouched (report fields are additive with defaults).
  • Docs: docs/explanation/layers-and-taxonomy.md gains the matrix vocabulary and the reachable-denominator rule; README documents hypothesis_matrix; the tutorial prints support_score.
  • Tests: tests/test_hypothesis_matrix.py (57 tests: taxonomy coverage, SUPPORTED ⟺ ladder-fires metamorphic equivalence, claim-floor truth table incl. non-finite beta_D and the A11 ensemble override, RFC-8259 serialisation with inf evidence, report-only non-influence, legacy default contract).

• Branch-shadowing report: ClassificationResult.triggered_hypotheses (issue #102 slice "assess branch shadowing", claim_status: pending). The classifier resolves by PRIORITY and returns exactly one dominant a_class, so a system that simultaneously shows, say, Mpemba overlap and pseudospectral phantom evidence reported only the first — the concurrently supported mechanism was silently erased. The new additive field reports every hypothesis that fires, in decision order, each as {rule_id, a_class, f_family, shadowed}. shadowed keys on the (class, family) pair, not on ladder position: several rungs can reach the same conclusion — gap_rate_consistency < 0.05 (with D17) fires the strong A1 rung and necessarily the residual < 0.20 rung too — and marking the second one shadowed would report a mechanism conflict where none exists. Every firing rung is still listed, so two rules supporting one conclusion stay visible as corroboration; only a genuinely different suppressed (class, family) counts as shadowing. The tuple is rule-level and deliberately not deduplicated — one suppressed mechanism can occupy several entries (A1 via both its rungs, A10/F5 via the pseudospectral and the M3a rung) — so the count of suppressed mechanisms is the number of distinct (a_class, f_family) pairs among the shadowed entries, not the number of entries. Documented with the counting recipe on the public README surface. To keep the report from drifting away from the decision it describes, the priority chain was extracted into a single _hypothesis_ladder() that evaluates all rungs and returns (rule_id, a_class, f_family, fires); both _pick_a_class() (first firing rung, else A12) and triggered_hypotheses() are derived from that one list, so there is no second copy of the conditions to fall out of sync. No verdict behaviour changed: class, family, verdict, tier and confidence are computed exactly as before (the ladder preserves the decision order; only the early returns became eager evaluation, and every directly indexed evidence key is unconditionally populated by _gather_evidence). The report is deliberately not consumed by any decision — letting it drive verdicts would be a classifier change and belongs behind the preregistered calibration study of issue #102. The confidence field is re-documented in _types.py as a heuristic support score, not a posterior probability; the confidence → support_score rename remains open. Serialised older results stay valid (default ()).

• D14 transient amplitude: projector baseline and norm geometry separated (issue #103, claim_status: pending). The legacy D14 trans_amplitude_ratio = sup_t ||e^{tL}||_2 conflated three questions. It is preserved byte-identically and re-documented precisely as an unstructured Hilbert-Schmidt semigroup norm estimate over the full complex Liouville space — not a state-amplitude ratio. Three additive, advisory fields split the confounds apart:

  • steady_projector(L) builds the asymptotic Riesz projector P_inf from an ordered Schur decomposition plus a Sylvester solve, not from a single arbitrary null vector, so a degenerate stationary manifold yields the correct rank-k conditional expectation. Semisimplicity of every peripheral mode is verified via rank deficiency of L - λI; a defective zero mode is fail-closed (semisimple=False, downstream value NaN). The peripheral tolerance is relative to ||L||_F, so the split is rate-unit invariant (consistent with issue #101).
  • D14b centered_transient_amplitude = sup_t ||e^{tL} − P_inf||_2, and D14d decaying_transient_amplitude = sup_t ||e^{tL}|_decay||_2 on the decaying invariant subspace in an orthonormal basis. These are not equivalent, contrary to the wording of issue #103: at t = 0 the centred form is ||I − P_inf||, and for a non-trivial oblique projector ||I − P|| = ||P||, so centring alone still carries exactly the baseline it was meant to remove. Only the restricted semigroup starts at 1. Both are reported so the difference stays visible; a regression test pins the identity.
  • D14c operational_trace_amplitude evaluates trace-norm amplification on traceless-Hermitian differences of density matrices, recorded as an explicit lower bound on the induced 1→1 norm (state family, seed and time grid are reported). For CPTP dynamics it must not exceed 1, which makes it its own contractivity control. No classifier change: the F2 branch keeps consuming the legacy D14 until the preregistered calibration study in issue #102.

  Hardened after cross-family review of PR #105: the peripheral cutoff is the Schur backward error n·eps·||L||_F, not sqrt(eps) (the latter absorbed genuinely resolved slow modes — a generator with rates 1 and 1e-8 reported rank 4 instead of 1); only genuinely zero modes count as stationary, and an oscillatory peripheral mode fails closed because e^{tL} has no time-independent limit then; the new diagnostics validate their time grids (finite, non-negative, strictly increasing — a backward-time grid evaluates the non-CPTP inverse and would fake a contractivity violation); their default grids include t = 0; the run seed is threaded into D14c and recorded in TransientResult.transient_seed; and compute_transient_layer computes the propagator sweep once and shares it across the variants.

• Scale-relative non-normality/pseudospectrum diagnostics (issue #101 slice A, claim_status: pending). One shared operator rate scale liouscope.numerics.scale.rate_scale(L) = ||L||_F (documented zero-operator semantics, fail-closed on non-finite input) now underpins additive, dimensionless variants of the rate-dimensioned legacy diagnostics: D8b henrici_relative = η_N/||L||_F in [0, 1] (clip-tolerance fail-closed), D10b kreiss_grid_lower_bound (dimensionless grid, local refinement, edge- maximizer + convergence metadata in the new KreissGridEstimate), scale- relative D11b/D12 (resolvent_peak_scaled, ridge_fwhm_rel), D13 with eps_abs = eps_rel · rate_scale (pseudospectral_radius_rel) and the new gap-directed intrusion diagnostic pseudospectral_abscissa(_rel) via numerics.pseudospec.pseudospectrum_extent (single-sweep radius+abscissa, NaN "under-resolved" marker instead of a fake 0.0). All are exactly invariant under a positive unit rescale L → cL for c ∈ {1e-10 … 1e10} — pinned by the new slice-B conformance suite tests/test_scale_conformance.py (invariance, rate-valued ~c scaling, D14 t → t/c metamorphic agreement, zero/normal/gapless-normal/Jordan oracles, unitary-basis invariance, fail-closed guards). The new fields are additive with NaN/False defaults on NonNormalityResult/ResolventResult (older callers and serialised results stay valid), surfaced as advisory evidence keys (ADVISORY_EVIDENCE_KEYS extended; the pinned metamorphic non-influence test covers them) and as pending-stamped D8b_henrici_relative/D10b_kreiss_scaled entries in the stability report. No classifier/verdict behaviour changed: per #101, the F5 gate switch is deferred to the preregistered calibration study + independent physics review. No manifest-contract change (run manifest fields unchanged, schema stays 1.5.0).

Changed

• Estimator labelling for D10 (issue #101 re-audit). The docstrings of kreiss_constant and the non-normality module no longer describe the legacy grid search as "Mitchell 2020": the value is a finite-grid lower bound without globality certificate. Values are byte-identical; docs only.
• Docs honesty (issues #101/#102 release policy). confidence is now documented as a deterministic heuristic support score (NOT calibrated; the tutorial's "calibrated 0..1" claim is fixed) and the docs state explicitly that the A10/F5 verdict path is not yet rate-unit invariant, with the new scale-relative diagnostics listed as pending advisory evidence (docs/explanation/layers-and-taxonomy.md).
• MANIFEST_SCHEMA_VERSION 1.4.0 → 1.5.0 — injective input-hash encoding (issue #97 item 4). compute_input_hash now absorbs each input object as a length-framed, type-tagged field (tag || len(payload) || payload) instead of a bare repr/byte concatenation. The old encoding was not injective: distinct input tuples could collide when their serialised forms concatenated to the same byte stream — e.g. compute_input_hash(12, 3) and compute_input_hash(1, 23) both hashed "123". Within diagnose() (fixed arity/types) the collision was practically unreachable, but compute_input_hash is exported public API, so the derivation is hardened and the schema stepped. Migration: input hashes and run IDs are, as always, comparable only within one schema_version; 1.4.0 manifests remain valid historical records but do not re-derive under 1.5.0. Pinned in tests/test_manifest.py (test_input_hash_framing_is_injective). Docs (README, docs/CANON_STATUS, docs/DEVELOPMENT_MIGRATION_0.6.0.dev0, reproducibility tutorial/how-to/ explanation) updated to 1.5.0.

Fixed

• steady_state / sparse_steady_state tolerance is now scale-relative (issue #97 item 5). The dense null-space tolerance was max(atol, n2·eps·s[0]) with a default absolute floor atol = 1e-9 in arbitrary rate units; a Liouvillian has rate dimension, so a pure change of units L → c·L flipped the uniqueness diagnosis. Symptom: 1e-10 · L for amplitude damping (unique steady state |0⟩⟨0|) raised DegenerateSteadyStateError with a wrong "null space has dimension 4" diagnosis, and with allow_degenerate=True returned a wrong state plus a warning asserting non-uniqueness as fact. The tolerance is now max(atol, max(rtol, n2·eps) · s[0]) with rtol = 1e-9 (relative, new keyword) and atol = 0.0 (absolute floor now opt-in); the near-zero-trace check inside the normaliser is decoupled from atol (the null-vector candidate has unit 2-norm, so that check is dimensionless). The sparse guard in sparse_steady_state gets the same semantics: guard threshold max(tol·scale, |sigma_shift|) with scale = sqrt(‖L‖₁‖L‖∞) (cheap upper bound on s_max) and a default sigma_shift chosen relative to that scale (None → 1e-8·scale) instead of the fixed absolute 1e-8. Fail-closed direction preserved: genuine degeneracy is detected at every scale. Pinned in tests/test_lindblad.py (…_invariant_under_rate_rescaling, …_degeneracy_still_detected_at_small_scale, …_atol_is_an_opt_in_absolute_floor) and tests/test_sparse.py (test_sparse_steady_state_diagnosis_invariant_under_rate_rescaling). API note: steady_state(..., atol=…) keeps its absolute-floor meaning but no longer doubles as the trace threshold; sparse_steady_state(..., sigma_shift=…) keeps its meaning when passed explicitly. Anchors unaffected (verified 2026-07-13: pytest tests/test_anchors.py green, all canonical fixtures are O(1)-scaled where relative ≈ old absolute tolerance).
• D7b entanglement_asymmetry now computes the Rylands charge-sector measure (issue #97 item 1). The previous implementation applied a full single-qubit Pauli twirl, which maximally mixes the twirled qubit and yields an entropy deficit S(rho_1) + ln2 − S(rho) rather than the published Ares–Murciano–Calabrese / Rylands entanglement asymmetry ΔS_A = S(Σ_q Π_q ρ Π_q) − S(ρ) with U(1) charge-sector projectors. The symptom: a Bell state (|01⟩+|10⟩)/√2, which lives entirely in the single charge sector q=1 and is therefore exactly symmetric (ΔS_A = 0), was reported as 2 ln2 ≈ 1.386. D7b is now the charge-block-dephasing measure (total magnetisation Q = n_1 + n_2 for the supported d=4 block). D7b is advisory-only — it never feeds a classifier verdict — and is not an anchor fixture, so this changes no tests/test_anchors.py behaviour and no gate outcome; it corrects a mislabelled report value. Pinned in tests/test_relaxation.py::test_d7b_entanglement_asymmetry_is_rylands_charge_measure. CITATION.cff stays pinned to released v0.5.0 per its own policy (the citable diagnostic surface — that D7b exists — is unchanged); the corrected methodology is recorded here for the next release's citation.
• AR(1) bias-correction docstrings corrected to match measured behaviour (issue #97 item 3, §6 Reality-Anchor). fitting/neff.py claimed the raw lag-1 estimator is biased by −(1+3ρ)/n and that the frozen first-order correction (ρ̂(n−1)+1)/(n−3) "removes the leading O(1/n) bias term". Monte-Carlo (40k reps) shows the raw bias is empirically closer to −(1+4ρ)/n and the correction leaves a residual O(1/n) downward bias of order −2ρ/n (e.g. −0.013 at n=80, ρ=0.5). The formula is unchanged — it stays the audit-frozen first-order closed form — but the docstrings now state the true cancelled/residual terms and tie the residual to the existing small-n warning, so the documented claim matches the code. No numerical/result change.
• Fail-closed hardening batch (2026-07-12 full-repo deep review). Three independent review passes (numerical core, classifier gates, IO/manifest/CI) found paths where malformed or non-finite input could silently upgrade a verdict, bypass a gate or fabricate a result. All fixes below are pinned in tests/test_failclosed_hardening.py (25 tests) plus additions to tests/test_sparse.py / tests/test_lindblad.py / tests/test_export.py; methodology-gated findings that need physics decisions are tracked in issue #97, not blind-fixed here.
  • A11 floor override now type-gated (diagnostics/api.py). Any duck-typed object exposing permits_claim_floor_override=True could lift the single-state maximally-mixed A11 floor, bypassing all provenance validation in ensemble.py; non-EnsembleEvidence input now raises TypeError fail-closed.
  • Non-finite symmetrised gaps no longer grant the issue-#80 A1 certificate (diagnostics/classification.py). kms_gap=inf (ratio collapses to 0.0 <= 1.2) or gns_gap=inf (passes the >= floor test) silently granted sym_gap_corroborated and upgraded A1 to CONFIRMED/PUBLICATION_GRADE 0.95; both certificate legs now require finite gaps.
  • F5 gapless phantom limit fires when the GNS gap is also floored (diagnostics/classification.py). With gap=0 AND gns_gap=0 (the realistic gapless case) the old inf > inf comparison was False and the documented gapless-phantom contract never fired, falling through to A12.
  • Malformed steady-state shapes fail closed in the maximally-mixed floor (diagnostics/classification.py). A still-vectorised (d^2,) steady state read as "not maximally mixed" and silently disabled the A11 floor; uninterpretable shapes now raise. The documented d < 2 placeholder behaviour is unchanged.
  • EnsembleEvidence rejects duplicate input_hashes (ensemble.py). Two paired runs with identical input hashes are the same system + initial state (no ordering-parameter variation) and are structurally not a reference-family comparison.
  • build_liouvillian / build_sparse_liouvillian input gates (core/lindblad.py, sparse/build.py). NaN/inf rates passed the sign test (NaN < 0 is False), an all-real ±inf diagonal H passed the Hermiticity gate, and the sparse builder skipped Hermiticity/rate/shape validation entirely (accepting non-GKSL generators the dense twin rejects). Both builders now share the same fail-closed gates; examples.v4_thermal_two_level additionally rejects non-finite beta/omega.
  • sparse_steady_state degeneracy guard (sparse/dense parity, S1 audit anchor) (sparse/arnoldi.py). Pure-dephasing-like degenerate NESS manifolds returned an arbitrary (not even PSD) ARPACK vector silently while the dense path raised DegenerateSteadyStateError; the sparse path now checks the two smallest-magnitude eigenvalues and fails closed, with the same allow_degenerate=True escape hatch.
  • steady_state no-null-vector fallback is no longer silent (core/lindblad.py). An invertible superoperator (no steady state at all) returned the smallest-singular-value direction with residual ||L rho|| = O(1) and no signal; the fallback now emits a RuntimeWarning carrying the residual, the docstring describes the actual mechanism (SVD direction, not "smallest-real-part eigenvector"), and degenerate (0, 0) / non-2-D inputs get structured ValueErrors instead of a bare IndexError.
  • diagnose() validates t_grid like every other boundary input (_diagnostics.py). A NaN-contaminated grid propagated through expm(L t) and still produced a finite, confident-looking beta_D; non-finite, negative, unordered or non-1-D grids now raise.
  • Zhou predictor early return honours caller-supplied arguments (_zhou.py). The all-zero-modes return hard-coded gap=0.0 / petermann_factor=nan, overwriting explicitly passed values so the manifest-grade record contradicted its own call.
  • dump_report writes RFC 8259-valid JSON (io/export.py). The default allow_nan=True emitted bare Infinity/NaN tokens (evidence ratios are legitimately inf), which strict consumers (JavaScript JSON.parse, Postgres jsonb, serde) reject; Python's lenient parser hid it from the round-trip tests. Non-finite floats are now tagged {"__nonfinite__": "inf" | "-inf" | "nan"} (mirroring the __complex__ tagging) and all three JSON writers (dump_report, dump_manifest, dump_stability_report) enforce allow_nan=False.
  • CI/tooling drift. The pre-commit ruff hook (v0.12.1) linted with a materially different rule engine than the CI gate (ruff==0.15.20) — now pinned together with a co-bump note. docs/requirements.txt was lower-bound-only while RTD builds with fail_on_warning: true; the RTD toolchain is now pinned exactly (sphinx 9.0.4 / furo 2025.12.19 / myst-parser 5.1.0, verified clean under sphinx-build -W).
• Contributor-docs drift. CONTRIBUTING.md still pointed the clone URL at the defunct coworker-research org (the same drift the packaging metadata fix already removed from pyproject.toml); it now points at marcohost33-maker/Liouscope. The lint command was aligned with AGENTS.md (ruff check src tests benchmarks — benchmarks/ is part of the CI lint gate and was missing from the contributor instructions).
• Support / governance statement in the README (issue #72 item 3, JOSS community gate). The Contributing section now states explicitly where to get help (GitHub issues, no separate channel) and who makes maintainer decisions (repository owner; methodology changes via CHANGELOG.md / ADRs).

Added

• Docs slice 2: Diátaxis sections written out (issue #72 item 2, follow-up slice). The Tutorials, How-to and Explanation stubs are replaced with hand-written content: two tutorials (first diagnostic run; reproducible runs with SPEC 7 rng and manifests), four how-to guides (manifest export/validation, QuTiP cross-checks, A11 EnsembleEvidence, D24 Zhou predictor) and three explanation pages (why "no single number", D1–D24 layers + A1–A12 taxonomy + verdict vocabulary, auditable reproducibility). All code snippets were executed against the current API as part of the change; sphinx-build -W stays clean (MyST dollarmath enabled for the math blocks).
• QuTiP dynamics differential oracle (tests/test_qutip_dynamics_oracle.py, 007acc portfolio-audit 2026-07-07 ask). The spectral oracle suite never exercised the time-domain propagation that the relaxation layer (D5–D7) fits against. New coverage: closed-form trajectory oracles for amplitude damping and pure dephasing, CPTP invariants (trace/Hermiticity/positivity) at every propagated time point, and QuTiP differentials — _evolve trajectories against qutip.mesolve (independent adaptive-ODE integrator) and expm(L t) against the exponential of QuTiP's independently built Liouvillian. 8 tests; the QuTiP half runs under the existing qutip marker / ci-qutip.yml job.
• FAIR / registry badges in the README (issue #72 item 5, "Kleinvieh"). PyPI version badge (the package is live on PyPI), Zenodo version-DOI badge (10.5281/zenodo.21246109, verified 2026-07-08) and the fair-software.eu badge at its honestly measured 4/5 (●●●●○: public repository, license, registry, citation; the checklist dot requires an OpenSSF Best Practices badge, which is future work). Badge state follows the howfairis checker semantics (Zenodo does not count as registry there — PyPI does). PEP 735 [dependency-groups] was evaluated and deliberately not added: per the issue's own criterion it only pays off once CI moves to uv, and mirroring the extras into groups today would create a second place for the dependency list to drift.
• Sphinx + Read the Docs documentation skeleton (issue #72 item 2). A docs/ Sphinx site (furo theme, MyST Markdown, autodoc + napoleon API reference) organised along the Diátaxis structure (tutorials / how-to / reference / explanation), plus a .readthedocs.yaml (v2) and docs/requirements.txt / a docs optional-dependency group. Builds clean under sphinx-build -W (warnings-as-errors); RTD mirrors that gate via fail_on_warning: true. The Reference section is auto-generated from the public liouscope.__all__ surface; tutorials / how-to / explanation are intentional stubs for a follow-up slice.
• JOSS paper skeleton (issue #72 item 3). paper/paper.md + paper/paper.bib following the JOSS 2026 section requirements (Summary, Statement of need, State of the field, Software design, Research impact, AI usage disclosure). All six bibliography DOIs are resolver-verified; author identities/ORCIDs and research-impact evidence are flagged TODO (not fabricated) and must be completed before submission.
• SPEC 7 rng keyword, additive phase (a) (issue #72 item 1). diagnose(), seed_everything() and the D18 surface (compute_lep_layer / initial_state_sensitivity) now accept a SPEC 7 rng argument (int, numpy.random.SeedSequence, Generator or BitGenerator) alongside the legacy seed; supplying both raises ValueError. rng inputs are normalised to a derived integer seed by the new public bridge liouscope.derive_seed (re-exported from liouscope.io.seed with the SeedLike/RNGLike aliases), and that derived value is what the run manifest records -- the manifest seed field, make_run_id derivation and schema version are unchanged, so a manifest alone still reproduces the run. Passing a Generator consumes one draw from the caller's generator (SPEC 7 consumption semantics). Legacy seed-only and no-argument calls remain byte-identical (defaults 42 / 7 preserved). seed stays fully supported in this phase; deprecation is a later, separate SPEC 7 phase. Contract pinned in tests/test_spec7_rng.py.

Changed

• Coverage ratchet 80 → 90 (issue #72 item 4). Measured branch coverage on the baseline (Python 3.14) was 94.61%. The --cov-fail-under gate in CI and the new [tool.coverage.report] fail_under in pyproject.toml are both set conservatively to 90 (up from the previous fixed 80), leaving headroom for cross-version variation on the 3.10–3.14 matrix. CONTRIBUTING.md updated to match.
• Release identity and provenance hardening. Released v0.5.0 remains immutable; default-branch VCS installs now report 0.6.0.dev0. Manifest schema 1.4.0 adds the canonical structured-ensemble-evidence digest to the input-hash domain when evidence participates; hashes are comparable only within one schema version.
• A11 typed evidence gate. A single state on rho_ss = I/d remains UNDEFINED / EXPLORATION. Public ensemble_confirmation=True is rejected fail-closed. Only validated immutable EnsembleEvidence with PASS and reason ENSEMBLE_MPEMBA_CONFIRMED may suppress the floor. The evidence binds manifest/run/input digests, family ordering, comparison and uncertainty methods, software version, and distinct producer/reviewer attestations; its digest enters the run hash and its payload remains in DiagnosticReport.extras.
• PyPI publication privilege-separated. Manual dispatch is build/QA only without OIDC. The upload job is restricted to a published release, explicit tag checkout, the protected pypi environment, the enable flag, and repeated source/artifact/tag/commit identity checks. External PyPI/Zenodo/SWHID gates in issue #50 remain open.

Added

• Classifier semantics debt B3/B4 made explicit contracts (issue #70, completes the sub-findings PR #81 left open; A5/A6/A8/A9 already on main). Both changes are behaviour-preserving -- no real-input classification result changes; the sacred anchor suite (tests/test_anchors.py) and the V1-V5 golden classification assertions stay byte-identical green.
  • B3 -- reserved A-classes A6/A7/A9. The taxonomy A1-A12-v3.1 defines twelve classes but _pick_a_class emits only nine; A6 (accelerated-decay), A7 (weak-dissipation singular, Mori 2024) and A9 (prethermalization/ETH) have no decision branch yet. They are now recorded in _consts.RESERVED_A_CLASSES as a discoverable code-level contract analogous to RESERVED_DIAGNOSTIC_SLOTS (D21-D23), so the "12 classes" name stays honest instead of silently unreachable. A static-AST reachability test forces the contract to stay in lock-step with the code (wiring A6/A7/A9 later must update the reserved set).
  • B4 -- advisory (unused) evidence named. lep_proximity (D16), bohr_ap_length (D11) and mpemba_expansion_alpha (D20) are surfaced in the evidence dict for audit but deliberately do NOT influence class/verdict/confidence. classification.ADVISORY_EVIDENCE_KEYS names that contract; a metamorphic test proves the non-influence (perturbing each key across {0, ±1e9, ±inf, nan} leaves the decision invariant). Wiring any of them (e.g. D18 initial_state_sensitivity as an A1 confidence dampener) is a class-influencing design decision with false-positive risk and is left to a dedicated PR with anchor + FP coverage -- not blind-hooked here. Pinned by tests/test_classifier_semantics_debt.py (8 tests, executed).
• Anti-overfit gates wired into the claim vocabulary (issue #71 B2). The residual-whiteness (Ljung-Box, fitting/whiteness.py) and temporal-holdout (fitting/holdout.py) gates were implemented and unit-tested but exported by nothing and connected to no claim-level concept. They are now exported from liouscope.fitting, and a new fitting/claim_gate.py::assess_relaxation_claim maps their verdicts onto the StabilityReport SAFE/REVIEW/BLOCK vocabulary fail-closed (worst-wins): a rejected holdout -> BLOCK, non-white residuals or an un-run gate -> REVIEW, both gates run and passing -> SAFE. The result feeds straight into build_stability_report(claim_level=...). This is purely additive -- diagnose() behaviour is unchanged; callers opt in. New tests/test_claim_gate.py pins the full truth table (10 tests, executed).
• D21-D23 reserved-slot contract (issue #71 C2). The D1-D24 schema name spans slots that are defined in the Drive-side canon but not implemented in this repository. _consts.RESERVED_DIAGNOSTIC_SLOTS now records D21-D23 as an explicit, discoverable code-level contract next to DIAGNOSTIC_SCHEMA_VERSION, each marked "reserved ... not implemented", so the "24 diagnostics" name stays honest. Pinned by tests/test_reserved_slots.py.

Changed

• MANIFEST_SCHEMA_VERSION bumped 1.2.0 -> 1.3.0 (provenance-derivation contract change; MIGRATION NOTE). The run_id/input_hash fix below widens the set of hashed inputs, so for the same physical input the derived input_hash and run_id VALUES now differ from v0.5.0/main (empirically: input_hash 1ac0e089… -> 20a2f552…, run_id face535e… -> 43016c65… on the reference small system). This is intentional (it fixes the collision described under Fixed), but it is backward-incompatible for provenance: a run_id archived under schema 1.2.0 (v0.5.0 or earlier) does NOT re-derive under 1.3.0 — the old key is stable and valid within 1.2.0, it is simply not comparable across the bump. The JSON manifest STRUCTURE (field list, types, additionalProperties:false) is unchanged, so a 1.2.0 manifest still parses; only the hash-derivation domain changed, which is exactly what the schema_version bump signals. Migration: consumers that key archives or reproducibility checks on run_id/input_hash must (a) treat hashes as comparable only within a single schema_version, and (b) re-emit manifests under 1.3.0 if a stable cross-version key is required — there is no value-level back-migration (the pre-bump inputs were never hashed). The MANIFEST_SCHEMA.json run_id/input_hash descriptions now state the schema_version-scoping explicitly so the constraint is machine-discoverable. (AGENTS.md Definition-of-Done #6: run-manifest contract touched -> schema bump + migration note.)
• D17 gap-rate consistency is now dimension-coherent (issue #69). D17 was |beta_D - Delta| / Delta, comparing the relative-entropy fit rate beta_D directly with the spectral gap Delta. Relative entropy near the steady state pi is quadratic in (rho - pi) when pi is faithful (full-rank), so it decays at 2*Delta; when pi is rank-deficient the null-space-leakage term is linear, so it decays at 1*Delta. That metric multiplier m in {1, 2} (verified empirically across V1-V5: m ~ 2.0 for faithful pi V1/V2/V4, m ~ 1.05 for rank-deficient pi V3/V5) inflated D17 (dephasing ~3.3, amp-damp ~1.1) and made the A1 "gap-controlled" label unreachable for every system. D17 now uses beta_D_linear, the dominant decay rate of the linear trace-distance curve (LIOU-F-018), which decays at the bare mode rate and is therefore dimension-coherent with Delta. The relative-entropy rate beta_D is unchanged (still the headline relaxation rate; the fit/bootstrap/anchor pipeline is untouched). The classifier now exposes beta_D, beta_D_linear, gap and the implied multiplier d17_metric_multiplier in the evidence dict (the factor is explicit and auditable, not hidden), and a genuine single-mode gap-controlled system can now earn A1 "gap-controlled" (family "none" -- see issue #70 A6 below, which corrected the A1 family from F1). The A1 early-branch is ordered after the F1-F5 gap-failure families and only before the relative-entropy-shape branches (M2/M3a/M3b): gap_rate_consistency + linear_fit_model are initial-state-dependent, so a strongly non-normal phantom/skin operator with an rho_0 that excites only the slow gap mode must NOT shadow its true (operator-intrinsic) F5/F2 mechanism -- it stays A10/F5 (resp. A4/F2), not A1 (Equalita #79 review). No V1-V5 mechanism label changes (golden-pinned end-to-end against baseline 3fef6a1); only the D17 number is corrected. New end-to-end regression tests/test_validation_systems/test_d17_gap_coherence.py (20 tests, incl. V1-V5 a_class golden + an adversarial non-normal-phantom shadow test) plus synthetic ordering regressions in tests/test_classification.py. Additive fields RelaxationResult.beta_D_linear/linear_fit_model and LepResult.beta_D_linear are defaulted, so serialised reports stay valid; compute_lep_layer(beta_D=...) -> compute_lep_layer(beta_D_linear=...) and the gap_rate_consistency(beta_D=...) param -> rate=... (both internal).
• U1 solver uncertainty now reads from a named nominal floor (issue #71 B5). compute_uncertainty_layer reported a bare 1e-10 magic number for U1 when no solver_residual was supplied; it now reads _consts.U1_NOMINAL_FLOOR and the docstring states plainly that U1 is a conservative placeholder (not a measured residual) unless the caller runs an ODE-tolerance sweep -- which diagnose() does not. Same numeric value, no behaviour change; the semantics are now explicit in code.
• Classifier semantics debt cleared: A1 family, F5 dimensional coherence, LEP degeneracy, dead EXCLUDED verdict (issue #70, A5/A6/A8/A9). Four semantic corrections to the A1-A12 classifier, each with a physics rationale:
  • A6 -- A1 now maps to family "none", not F1. F1 is the Mori-Shirai overlap gap-FAILURE mechanism (PRL 125, 230604); A1 (asymptotic-gap- controlled, primitive QMS) is precisely the no-gap-failure case. Tagging the healthy gap-controlled label with a gap-FAILURE family was a category error. "none" = "No gap-failure mechanism flagged" is A1's correct family. (tests/test_classification.py + test_d17_gap_coherence.py updated to pin A1/"none"; the #69 dimension-coherence logic is untouched.)
  • A8 -- the F5 phantom-relaxation rule is now dimension-coherent and scale-invariant. The old rule pseudospectral_radius > 2 * gap_to_gns_ratio compared a RATE (the D13 radius max{|z|: z in sigma_eps(L)}) against a dimensionless ratio, so rescaling the Liouvillian L -> cL (a pure change of time unit) flipped the A10/F5 verdict. The rule now compares the dimensionless pseudospectral reach radius / gap (how far the eps-pseudospectrum extends relative to the asymptotic decay rate Delta, the physical phantom signature, Znidaric 2023) against 2 * gap_to_gns_ratio -- both sides dimensionless, scale-invariant to leading order. A vanishing gap is treated as infinite reach (the gapless/critical phantom limit). New metamorphic rescale tests over c in [1e-3, 1e3] pin no verdict flip, plus a regression proving the old bare-radius rule would have flipped.
  • A9 -- lep_proximity (D16) no longer blind to exact degeneracies. The min-separation scan skipped every pair with sep <= atol, so an exactly degenerate eigenvalue pair -- the STRONGEST exceptional-point signal (two eigenvalues coalescing) -- was invisible, and a fully degenerate spectrum returned inf ("maximally far from an EP"), the exact inverse of the physics. Coalescence now yields proximity 0.0; the candidate-count loop uses the same data and a max(10*min_sep, atol) window, so the two loops are mutually consistent. (D16 measures eigenvalue proximity only; genuine defective EP vs semisimple degeneracy is disambiguated by the Petermann factor D9, as before.) New edge tests cover exact/full/sub-atol degeneracy.
  • A5 -- the unreachable EXCLUDED verdict was removed from the Verdict Literal (and _consts.VERDICT_EXCLUDED). A single-pass, maximum-evidence classifier reports the best-fit A-class with its support and never reports a class it is simultaneously ruling out, so a per-class active-rejection verdict is not expressible in this architecture -- the value was permanently unreachable through diagnose(). The old confidence < 0.30 -> EXCLUDED branch was also semantically wrong (low confidence = epistemic "unresolved" = NOT_EXCLUDED, not counter-evidence); low confidence now correctly yields NOT_EXCLUDED. API note: Verdict narrows from 5 to 4 members; this is a type-surface narrowing only (runtime diagnose() output is unchanged -- it never emitted EXCLUDED). Genuine active exclusion is deferred to a future per-hypothesis scoring mode. B3 (unreachable classes A6/A7/A9) and B4 (unused D16/D18/D11/D12/D20 evidence) are analysed but NOT wired in this PR (see PR body) -- both are class/verdict-influencing design decisions that can introduce false positives and are deferred to dedicated PRs with anchor coverage.

Fixed

• A1 PUBLICATION_GRADE now requires a positive symmetrised-gap certificate, not threshold exhaustiveness (issue #80; classification-semantics change in a dedicated PR per AGENTS.md §3; the twin of the #88 fix). The A1 early branch awarded CONFIRMED/PUBLICATION_GRADE (confidence 0.95) whenever gap_rate_consistency < 0.05 + single-exp held and none of the F1-F5 thresholds fired — so the publication-grade claim rested on the unprovable exhaustiveness of the F1-F5 threshold set (both #79 reviewers' residual): a hypothetical weakly-non-normal gap failure below all thresholds with a single-exp-at-gap trajectory would self-certify. The burden of proof is now reversed (issue #80 to-do 2): a new evidence key sym_gap_corroborated (1.0 iff a measured symmetrised gap shows no F3-grade reduction — certified GNS gap_to_gns_ratio <= 1.2, or KMS gap_to_kms_ratio <= 1.2) gates the 0.95 score; uncorroborated A1 caps at 0.70 → CANDIDATE/CONFIRMATION (honest grade: the single-exp-at-gap observable is measured, operator-intrinsic gap control is not). gap_to_kms_ratio thereby graduates from advisory (#89) to class-influencing; using it as an F3 veto stays deferred. Anchor-preserving: the gap-controlled thermal reference has Delta_GNS = Delta_KMS = Delta exactly (corroborated → 0.95 unchanged); V1-V5 golden labels are A5/A11 and never take the A1 confidence path; the sacred anchor suite stays green. Synthetic adversary + KMS-certificate + double-floored fail-closed tests added. claim_status: pending until cross-family review confirms the semantics.
• A2/F3 no longer fires off the gns_gap conservative-floor sentinel (issue #88; classification-semantics change in a dedicated PR per AGENTS.md §3). diagnostics.spectral.gns_gap deliberately floors Delta_GNS to ~0 when the GNS symmetrisation certifies no contraction (the documented 2026-07 audit-A1 behaviour for non-detailed-balance steady states carrying coherences; that floor is unchanged). But _gather_evidence turned the sentinel into gap_to_gns_ratio ~ 1e10..inf, and the F3 branch plus _confidence promoted the exploded ratio straight to A2/F3 CONFIRMED / PUBLICATION_GRADE — a publication-grade Mori-Shirai-2023 mechanism claim keyed on the absence of a certificate, on inputs as tame as a textbook Rabi-driven amplitude-damped qubit with Delta_KMS == Delta (provably no real symmetrised-gap reduction). Fix (issue #88 option 1: positive-evidence burden): the evidence dict now carries gns_certified (1.0 iff gns_gap >= GNS_CERTIFIED_RTOL * gap, new named constant 1e-8 in _consts), and both the F3 branch and the A2 high-confidence rule require it — the uncertified sentinel falls through to the state-dependent shape branches (honest floor: absence of evidence). gap_to_kms_ratio is now surfaced as advisory audit context (issue #88 option 2 — wiring it as an F3 veto is deferred to its own FP study); the advisory metamorphic contract covers it. Verdict flips (uncovered input class only): the #88 repro family (Rabi qubit, drives 0.3/0.7/1.5) flips A2/F3 CONFIRMED/ PUBLICATION_GRADE (0.85) → A8 or A10-via-M3a CANDIDATE/CONFIRMATION (≤0.70). Anchor-preserving: V1/V3/V4 have diagonal steady states (finite, certified gns_gap), V5 is caught by the Mpemba branch first; the sacred anchor suite and all V1-V5 golden assertions stay green. A certified reduction still fires A2/F3 at 0.85 (positive-control test). New e2e guard tests/test_classifier_f3_sentinel.py; claim_status: pending until cross-family review confirms the semantics.
• Hermiticity / normalisation validation gates are now absolute, not ~10^4x looser than advertised. numerics.linalg.is_hermitian (and therefore is_density_matrix), the Hamiltonian Hermiticity check in core.lindblad.build_liouvillian ("H must be Hermitian within 1e-9 atol"), and the unit-norm check in core.jumps.engineered_target_jumps all called np.allclose/np.isclose with only atol= set. NumPy's default rtol=1e-5 silently widened each gate to atol + 1e-5*|entry| (~1e-5 for O(1) density matrices / Hamiltonians), so a matrix non-Hermitian at 1e-6 — a thousand times the documented 1e-9 tolerance — passed as valid. All three now pass rtol=0.0 so the atol means exactly what the docstring/error says. Behaviour-preserving for every physical operator (Hermitian to machine precision); only genuinely malformed input near the old blind spot is now rejected. New regression tests/test_numerics.py:: test_is_hermitian_atol_is_absolute_not_relative.
• GLS AR(1) log-likelihood is now the exact Prais-Winsten likelihood. fitting.gls.fit_gls_ar1 whitens with the exact AR(1) transform (keeping observation 0 scaled by sqrt(1-rho^2), all n points), but reported gaussian_log_likelihood(whitened) — the iid-Gaussian likelihood of the whitened residuals, which omits the transform's log-Jacobian +0.5*log(1-rho^2). The reported value was therefore a hybrid of the exact and conditional likelihoods. Because each model M0..M3b fits its own rho and its per-model log_likelihood feeds aicc()/choose_model (which drives mechanism classification), the missing rho-dependent term biased cross-model AICc toward under-fitting high-rho fits. The Jacobian is now added. Anchor-preserving: the sacred tests/test_anchors.py suite and the V1-V5 golden classification assertions stay green (the correction does not flip the selected model on any canonical fixture); it only affects near-tied AICc comparisons on other data. No manifest-contract or citable-claim change (the run manifest does not record per-model likelihoods; input_hash is derived from inputs, not outputs). New regression tests/test_fitting.py:: test_gls_ar1_log_likelihood_is_exact_prais_winsten.
• run_id / input_hash now cover every output-affecting input. Both provenance keys were derived only from (L_super, rho_initial) (plus seed and framework_version for run_id), so two diagnose() calls that differed only in an analysis knob — bootstrap_B, t_grid, include_mpemba, solver_path, or an explicitly-supplied rho_steady_state — produced byte-different reports carrying an identical run_id. That is a collision: the MANIFEST_SCHEMA documents run_id as "deterministically derived from input parameters", but a larger bootstrap_B (wider BCa CI) reused the same key, breaking archival keyed on run_id. compute_input_hash now folds all output-affecting arguments; the MANIFEST_SCHEMA run_id/input_hash descriptions are made accurate. Determinism for identical inputs is unchanged (repeated runs still collide, as required). New regression tests/test_manifest.py::test_run_id_distinguishes_analysis_config.
• seed_everything fails closed on out-of-range seeds before mutating any global state. Legacy np.random.seed only accepts 0 <= seed < 2**32; a seed of 2**32 passed the previous non-negative-int guard, then raised mid-function after PYTHONHASHSEED and random.seed had already been changed (partial, inconsistent state). The bound is now checked up front. New regression in tests/test_input_guards.py.
• D14/D15 time grid stays strictly ascending under extreme non-normality. _physics_time_grid built its log-spaced early segment as geomspace(t_decay * 1e-6, t_early, …); when the numerical abscissa exceeds the gap by more than six decades the start overtakes t_early and the segment ran backwards, clustering the fine sampling after the transient peak instead of before it. The start is now clamped below t_early in that regime (a no-op for all normal spectra). New regression in tests/test_transient.py.
• D13 pseudospectral_radius fails closed on non-finite operators. The grid search fed L straight into eigvals/svdvals with no finite/square guard, unlike the rest of numerics (linalg.py, cptp.py); a NaN/inf operator surfaced as an opaque LAPACK error deep in the svd loop instead of a located, argument-named ValueError. Added require_finite_square_2d at the entry point. New regression in tests/test_resolvent.py.
• Metadata / provenance consistency. codemeta.json no longer overclaims "D21-D24 implemented" (it now matches CITATION.cff: D1-D20 + D2b/D7b/D11b + the opt-in D24, with D21-D23 as reserved schema slots); MANIFEST_SCHEMA.json $id points at the canonical marcohost33-maker/Liouscope repo instead of a stale org; and build_stability_report casts cp_choi_min_eig through float() like every other numeric it stores, so a np.float32/np.longdouble input cannot break JSON serialization.
• Removed a dead classifier branch. The A1 gap_rate_consistency < 0.05 and aicc_model == "M0" pre-check returned the identical ("A1", "none") that the following < 0.20 line already returns, and the A1 confidence keys on gap_rate_consistency alone — so it changed neither label nor score. Deleting it removes an implied distinction that does not exist (no behaviour change).
• D16 lep_proximity now fails closed on non-finite eigenvalues (issue #82, part 2). NaN / +-inf eigenvalues were silently swallowed by the pairwise comparisons and yielded a finite (proximity, count) result -- e.g. a NaN input returned (1.0, 1), a spurious LEP classification instead of an error. lep_proximity (and therefore compute_lep_layer) now raises ValueError, listing the offending indices, before the scan. This closes a Silent-Failure- Gate violation; the exact/full/sub-atol degeneracy behaviour is unchanged. Pinned by new negative-path tests in tests/test_lep.py.
• A8/F5 scale-invariance test no longer overclaims (issue #82, part 1). The test test_a8_f5_rule_is_scale_invariant_under_rescale hand-scaled both the gap and the pseudospectral radius by the same factor, making radius/gap trivially constant, yet its name claimed full scale-invariance. It is renamed to test_a8_f5_decision_rule_invariant_under_exactly_scaled_evidence (it pins the decision rule on idealised evidence only) and a new end-to-end metamorphic test recomputes the D13 pseudospectral radius from a genuinely rescaled operator c*M. Because D13 uses a fixed eps = 1e-3, the reach radius/gap is scale-invariant to leading order only (measured drift ~1.34x over six decades, converging to the true spectral-radius/gap at large scale); the classifier docstring already said "to leading order" -- only the test/PR wording overclaimed. Relative-eps rescaling for exact invariance is a physics-design follow-up, out of scope here.
• D19/A11 Mpemba detector no longer false-positives on trivially symmetric initial states (issue #68). Three coupled defects were corrected:
  • diagnostics/mpemba.py::overlap_c1 normalised the slowest-mode overlap by ||l_1||, so only its zero test was meaningful. It now returns the true biorthogonal expansion coefficient |<l_1, rho_0>| / |<l_1, r_1>| (denominator floored at EPS_DIV), the physical weight of the slow mode.
  • A non-triviality guard (is_trivial_overlap, wired through diagnose() via the new rho_steady_state argument of compute_mpemba_layer) now distinguishes a symmetry-protected zero overlap from an anomalous skip. In the sector decomposition set by the eigenprojectors of rho_ss (robust to a degenerate / maximally mixed steady state, unlike an eigenvector basis), a rho_0 whose active blocks are disjoint from the slowest mode's can never populate it — trivially fast relaxation, not Mpemba. Amplitude damping with the default rho_0 = I/2 (and the excited state |1><1|) previously returned A11 F4 CONFIRMED PUBLICATION_GRADE; both now fall through the A11 branch. MpembaResult gains a trivial_overlap field.
  • A single initial state skipping the slowest mode is now CANDIDATE-grade, not PUBLICATION_GRADE: A11 confidence is capped at 0.70 because confirmation needs a reference family (e.g. thermal states across temperatures) the single-state pipeline does not provide. The README dephasing example (rho_ss = I/2, which collapses to a single eigenprojector so the triviality guard cannot fire) stays A11 but as CANDIDATE/CONFIRMATION, no longer self-certifying.
  • New/updated coverage in tests/test_mpemba.py (biorthogonal coefficient, guard truth table, a fine-tuned qutrit true-positive oracle, canonical false-positive regressions) and tests/test_classification.py.
• CI: pinned ruff==0.15.20 and reformatted the _diagnostics import block. Ruff's isort I001 heuristic for lines-after-imports before a module-level constant changed between 0.15.14 and 0.15.20, flagging a byte-identical, previously-green import block and turning the whole test matrix red on a tool release rather than a code defect. Ruff is now pinned exactly in both pyproject (dev extra) and the reusable ci-python-local.yml install step; a bump is now a deliberate, reformat-in-one-PR event.
• diagnose() fails closed on the solver path. An unknown solver_path raises ValueError; the reserved sparse_arpack path raises NotImplementedError instead of silently falling back to the dense solver. Now covered by tests/test_solver_path_contract.py (2 tests, executed).
• D2/D2b symmetrised gaps (numerics.adjoint, diagnostics.spectral) are now genuine Gram-adjoint constructions (2026-07 audit A1). Three coupled defects were fixed:
  • alicki_adjoint built the pi-weighted adjoint as (rho^-1 (x) I) L (rho (x) I) while the GNS similarity transform used the anchor-B Gram G = rho.T (x) I. The two coincide only for real-diagonal steady states; off that manifold the "symmetrised" generator was not G-Hermitian, its silent Hermitisation masked the defect, and gns_gap() returned unphysical negative values (repro: generic d=3 GKSL with steady-state coherences gave Delta_s = -4.4 at Delta = 0.63). The Kron factors are now transposed — i.e. exactly G^{-1} L G. No-op for the real-diagonal anchor fixtures (all 21 anchors unchanged).
  • kms_gap mixed pictures: it symmetrised with the GNS adjoint but conjugated with the KMS Gram. It now uses the KMS-Gram adjoint via the new numerics.adjoint.gram_adjoint(L, G) helper.
  • Gap extraction (_real_gap_from_symmetric) now deflates exactly ONE steady-state zero mode instead of filtering all near-zero eigenvalues. A degenerate Hermitian-part kernel means there is no certified exponential GNS contraction (Delta_s = 0); the old filtering over-certified the bound (Delta_s > Delta, contradicting the Kadison-Schwarz contraction property). Consequence: driven pure dephasing with rho_ss = I/2 now honestly reports gns_gap = 0 (the sigma_z mode has zero instantaneous GNS decay rate); detailed-balance systems keep Delta_s = Delta. New contraction-bound regression tests (0 <= Delta_s <= Delta, 0 <= Delta_KMS <= Delta) pin the fix on non-diagonal steady states (40/40 random systems verified).
• sparse.chi1.chi1_lower_bound returned K^(1/4) instead of the documented K^(1/2) (2026-07 audit A4): the ratio ||r||*||l|| / |<l,r>| is already the square root of the Petermann factor; the code took a second square root, quadratically weakening the ARPACK-reliability certificate. A regression test now pins the certificate against the dense petermann_factors oracle (the old test only asserted chi > 0).
• Silent zero-width confidence interval on bootstrap failure (diagnostics.relaxation.compute_relaxation_layer, 2026-07 audit A7): a swallowed parametric_bootstrap/bca_ci exception used to keep the degenerate initialisation (beta_D, beta_D), which the uncertainty layer read as fit_uncertainty = 0.0 — perfect certainty as the failure mode. The CI is now (nan, nan) plus a RuntimeWarning; compute_uncertainty_layer already maps that to fit_uncertainty = nan.
• Robustness guards (2026-07 audit A10/A11): core.lindblad.steady_state casts integer/bool input instead of crashing in np.finfo; examples.v4_thermal_two_level rejects beta*omega <= 0 (divide-by-zero); io.seed.seed_everything rejects bool seeds (silent 0/1 seeding); core.jumps.engineered_target_jumps rejects a zero target vector and its docstring now matches the two-operator return.
• Docstring/comment drift fixed against implementations: D4 spread comment in _types.py, pseudospectrum grid location, D18 sensitivity ensemble, D7b fallback value, diagnostics/__init__ layer count.
• liouscope.numerics.cptp: fail-closed input validation now rejects non-finite generator/channel entries (L_super, channel_super) and invalid (negative / non-finite) tolerances before the matrix exponential, Choi construction, or eigensolvers run. This replaces opaque LAPACK failures with a clear ValueError and prevents a bad caller tolerance from silently inverting the CP/TP verdict.

Added

• Property-based test suite (tests/test_property_based.py, Hypothesis): ground-truth-free invariants on randomly drawn GKSL systems — gap scaling gap(c*L) = c*gap(L), spectrum invariance under unitary similarity, trace-distance axioms + CPTP contractivity, the GKSL semigroup law Phi_{t1+t2} = Phi_{t2} Phi_{t1}, and no-false-alarm coverage for the Choi CPTP gate. derandomize=True keeps CI deterministic (the hypothesis dev dependency was previously declared but unused).
• CodeQL SAST workflow (.github/workflows/codeql.yml): weekly + PR/push security-and-quality Python analysis, SHA-pinned, minimal permissions (OpenSSF Scorecard SAST check; ruff is a linter, not SAST).
• pre-commit configuration (.pre-commit-config.yaml): ruff-check, zizmor (regular persona), CITATION.cff validation (cffconvert) and standard hygiene hooks. No formatter (history churn) — CI remains the authoritative gate.
• numerics.adjoint.gram_adjoint(L_super, G): adjoint of the Heisenberg-picture generator w.r.t. an arbitrary Gram matrix (GNS and KMS are the two instantiations).
• jsonschema>=4.18 added to the dev extra so CI actually exercises the full Draft202012Validator structural-conformance path for run manifests. It was previously absent from the dev/CI environment, so _compiled_validator() returned None and the manifest schema-contract test was silently skipped on every CI run (the built-in required-field fallback still ran). No runtime change — jsonschema remains an optional runtime extra with graceful fallback.
• tests/test_cptp_choi.py: regression coverage for the new non-finite / invalid tolerance seams, plus an independent partial-trace oracle that pins the choi_matrix tensor-leg convention (the identity channel's Choi matrix is the unnormalised maximally entangled operator, not eye(d**2)) and its CP-boundary PSD-ness across d in {2, 4, 8, 16}.

CI

• ci.yml/ci-qutip.yml: concurrency groups auto-cancel superseded runs on non-main refs (sp-repo-review GH102).
• pypi.yml: fail-closed release QA gate (twine check --strict + check-wheel-contents) between build and Trusted-Publishing upload.
• scorecard.yml: publish_results: true now that the repo is public (badge-capable, externally auditable score); stale private-repo comment refreshed.
• pyproject.toml strictness batch (sp-repo-review): pytest gains --strict-config, xfail_strict, log_level and filterwarnings = ["error", ...] (unexpected warnings — including deprecations from dependency drift — now fail CI; the documented ResolventConvergenceWarning lower-bound path is allowlisted); mypy gains warn_unreachable + enable_error_code = [ignore-without-code, redundant-expr, truthy-bool]; wheel removed from build-system.requires (setuptools adds it itself when needed); redundant ruff target-version dropped (inferred from requires-python).
• pypi.yml: the Trusted-Publishing step now sets print-hash: true so each uploaded sdist/wheel's SHA-256 is logged for the release-evidence lock. No second attestation step is added — the PyPA action already uploads a PEP 740 attestation by default under Trusted Publishing, and PyPI rejects duplicate predicates / more than two attestations per file. docs/RELEASE_AUDIT_v0.5.0.md §5 documents the evidence flow; the stale "private repo" comment was refreshed to "public".

Documentation

• README: the diagnostic layer table now uses the code-backed D-numbering (module docstrings / StabilityReport keys as source of truth) — the previous table circulated a second, contradictory numbering scheme. The "24 diagnostics" claim is qualified honestly: D1-D20 (+D2b/D7b/D11b) and D24 are code-backed; D21-D23 are schema-defined but not yet implemented.
• README/AGENTS.md reproducibility claims corrected to what the code does: the manifest does not (yet) record lattice geometry / dissipator family / full result graph; seed_everything does not control SciPy/BLAS threading.
• README quickstart no longer suggests the dephased-qubit example classifies as "A1" (see the tracked Mpemba false-positive issue).
• Synchronized docs/CANON_STATUS.md and AGENTS.md with the v0.5.0 runtime canon: public-repo status, Python 3.14 CI coverage, the dedicated QuTiP cross-check checks, and the StabilityReport v2.1 additive projection.
• Added an Architecture Decision Record directory (docs/adr/) with ADR 0001 — Scientific-Python support policy (SPEC 0). ADR 0001 records the decision to follow SPEC 0 and the post-v0.5 target (requires-python >=3.12, CI 3.12/3.13/3.14, drop 3.10/3.11 in a dedicated support-policy release). It is a decision record only — pyproject.toml, the CI matrix, and classifiers are unchanged in this PR.
• Consolidated the duplicate support-policy ADR: the governance detail from docs/ADR_SUPPORT_POLICY.md (release classification, evidence gate, the required-steps checklist) is merged into docs/adr/0001-python-support-policy.md, and the old path is reduced to a pointer stub so existing links still resolve. Single numbered ADR going forward.

0.5.0 - 2026-06-25

Release cut: the Canon v0.5 diagnostics & contracts wave plus the cross-family CPTP-Choi hardening (PR #55, B1/B2) and the repo-wide mypy gate fix that shipped on main after v0.4.1. MINOR bump per SemVer: backward-compatible, additive API surface (8 validated formelbuch entries LIOU-A-011/A-012/A-013, F-018/F-019/F-020, RPT-001, NG-003; new StabilityReport v2.1 projection and cptp Choi gate). No diagnose() / DiagnosticReport break, no manifest schema bump (the new StabilityReport is a separate, additive projection).

Fixed

• CPTP Choi gate (liouscope.numerics.cptp) hardened against non-GKSL / corrupted input (cross-family math review of PR #55, B1/B2). The gate now fail-closes on two seams a naive Choi test waved through:
  • B1 — trace preservation is checked at the propagator, not the generator. The TP residual was ||<<I| M_L|| (the generator), which is dt-/scale-blind: a sub-tolerance generator violation amplified by a large dt is a gross propagator trace violation that an absolute-tolerance generator check passes. Repro: generator residual 7.07e-10 (< the old 1e-9) but propagator trace scaled ~148x -> real TP residual ~208; the old gate reported is_tp=True. Now TP is ||<<I| Phi - <<I|| on Phi = exp(dt*L) -> is_tp=False.
  • B2 — Choi Hermiticity is checked before the PSD test. A non-Hermiticity- preserving map has a non-Hermitian Choi matrix; Hermitising it ((J+J^dag)/2) before taking the minimum eigenvalue masked the defect. Repro: a depolarizing channel plus a small non-HP term has Hermitised min_eig > 0 (old PSD check "CP") yet ||J - J^dag|| > 0; the old gate reported is_cp=True. Now a non-Hermitian Choi forces is_cp=False.
  • Absolute 1e-9 tolerances replaced by relative ones (scaled by ||J|| / sqrt(d)) so the verdict is scale-invariant. ChoiGateResult gains choi_herm_residual and is_hp. The gate's CP claim therefore now also fail-closes on non-GKSL / corrupted input, not only on physical channels. Regression tests added for both repros in tests/test_cptp_choi.py.
• liouscope.fitting.holdout.train_holdout_split now rejects a non-strictly- increasing (or non-finite) time grid, preventing future-sample leakage into training from an unordered series (cross-family review c3).
• liouscope.fitting.whiteness docstring: precise Ljung-Box (1978) citation (Biometrika 65(2), 297-303) and an explicit statement of the dof = m - n_params choice (default n_params=0, the conservative fail-closed dof; cross-family review c2). No behaviour change.
• Repo-wide mypy gate failure on the Python 3.12-3.14 CI matrix. NumPy >=2.5 ships type stubs that use the PEP 695 type statement, which mypy rejects while parsing numpy/__init__.pyi unless its target is >=3.12. Raised [tool.mypy] python_version from 3.10 to 3.12. Runtime 3.10/3.11 support is still guarded by the test matrix and ruff target-version = py310; the change only affects type-checking semantics, not packaged code. (Preferred over capping numpy<2.5 so v0.5 development stays on the current NumPy.)

Added

• Canon v0.5 diagnostics & contracts (additive, backward-compatible; validated formelbuch entries LIOU-A-011/A-012/A-013, F-018/F-019/F-020, RPT-001, NG-003). Each new diagnostic is pinned to an independent oracle (closed form, QuTiP, or analytic soll-value), never to its own machinery:

  • CPTP Choi-PSD gate (liouscope.numerics.cptp, LIOU-A-011): verifies complete positivity of exp(dt*L) via the Choi-matrix minimum eigenvalue (>= -tol) plus the trace-preservation residual || <<I| M_L ||. Oracles: the transpose map (positive but not CP) has Choi min_eig = -1; a dephasing channel sits on the CP boundary at min_eig = 0 (entry beleg); and an Euler step I + dt*L is shown non-CP where exp(dt*L) stays CP (entry NR-002: Euler positivity is not a CP proof). Dense-only CP proof by design.
  • Trace distance D_tr (liouscope.diagnostics.relaxation.trace_distance, LIOU-F-018): (1/2)||rho-sigma||_1, the observable relaxation metric beside D5/D6/D7, plus an additive RelaxationResult.trace_distance_curve. Oracles: orthogonal pure states -> 1, diagonal states -> total variation, Fuchs-van de Graaf bounds vs the Uhlmann fidelity, CPTP contractivity, and qutip.tracedist.
  • Temporal holdout split (liouscope.fitting.holdout, LIOU-A-012): an out-of-sample anti-overfit gate for the M0-M3b hierarchy (time-ordered tail, no shuffling).
  • Residual-whiteness gate (liouscope.fitting.whiteness, LIOU-A-013): Ljung-Box Q with the chi-squared reference from scipy.stats (white noise passes, AR(1) is rejected).
  • Metamorphic spectral oracles (tests/test_metamorphic_spectral.py, LIOU-F-020): gap(c*L)=c*gap(L) and spec(U L U^dag)=spec(L) — ground-truth- free invariants.
  • Gap-invariant reproduction (tests/test_gap_invariants_canon.py, LIOU-F-019): reproduces the pack mini-oracle parametrisation (amplitude damping gap=gamma/2; thermal g_down=0.9, g_up=0.2, omega=1.3 -> gap=0.55). NOTE: the general gamma/2 and (g_up+g_down)/2 oracles already exist in tests/test_qutip_spectral_oracle.py (PR#51); this adds the exact pack parametrisation. Dephasing 2*gamma is intentionally NOT re-added (canon).
  • StabilityReport v2.1 contract (liouscope.io.stability_report + packaged STABILITY_REPORT_SCHEMA.json, LIOU-RPT-001): a claim-safe, machine-auditable projection of a DiagnosticReport adding claim_level (SAFE/REVIEW/BLOCK), direction, cp_evidence_level, independently recomputed invariant residuals, an evidence_bundle and provenance. New diagnostics carry claim_status="pending". Purely additive — it does NOT modify the existing run manifest, MANIFEST_SCHEMA.json or DiagnosticReport (no schema-version bump; older artefacts stay valid).
  • Petermann interpretation caveat (D9 docstring, LIOU-NG-003): a large Petermann factor is necessary-but-not-sufficient for transient amplification; sup_t ||e^{tL}|| is bounded by the Kreiss constant (D10) / numerical abscissa (D15), not by the Petermann factor alone.

  Every new public boundary ships negative/edge-input gates (negative dt, NaN, non-square dim, bad holdout_frac, m>=N, non-positive dof, out-of-enum verdict fields, tampered schema fields). Verified locally on CPython 3.14 via the CI command chain: ruff check src tests benchmarks (exit 0), mypy src/liouscope clean at --python-version 3.12 (the pyproject-default invocation aborts on an unrelated numpy-stub/toolchain skew in the local sandbox), full suite pytest --cov-fail-under=80 (375 passed, coverage 93.9%), anchors pytest tests/test_anchors.py (21 passed), pytest -m qutip (8 passed).

• Independent-oracle cross-checks for the non-normality layer D8-D11 (tests/test_nonnormality_oracle.py). The previous tests (tests/test_nonnormality.py) only asserted signs and self-consistency (eta > 0, K > 0, length >= 1), so a wrong normalisation or eigenvalue filter would pass. The new module pins each diagnostic to a closed-form or an independent numerical oracle, never to the library's own machinery: D8 Henrici against the unitarily-invariant sqrt(||A||_F^2 - sum|lambda|^2) (Henrici 1962; closed form |b| for [[0,b],[0,-1]]); D9 Petermann against the 2x2 adjugate condition-number tr(B0^H B0)/|li-lj|^2 (e.g. Phys. Rev. Research 5, 033042 (2023)) plus the normality floor (Petermann = 1 for a pure-dephasing Liouvillian); D10 Kreiss against the continuous-time reference facts (K = 1 for a normal Hurwitz matrix; K >= 1 always; closed form sqrt(1+b^2) for [[0,b],[0,-1]]) and an independent 2D Nelder-Mead resolvent-norm maximisation (Kreiss matrix theorem; Mitchell, SIAM J. Matrix Anal. Appl. 41(4), 2020); D11 Bohr AP against hand-constructed spectra with a known longest arithmetic progression and the log_2(d) Pauli bound (Basso, arXiv:2510.07267, 2025). Each oracle includes a non-vacuous negative control (wrong-zero Henrici, missing-denominator Petermann via eigenvalue gap != 1, grid-cannot-exceed-oracle Kreiss, no-three-term-AP Bohr). Test-only; no production code, anchor, or DiagnosticReport output changes. Verified locally via the CI command chain on CPython 3.14: ruff check src tests (exit 0), mypy src/liouscope (exit 0), pytest -q full suite (307 passed, 1 skipped — pre-existing local jsonschema-extra skip), pytest -m qutip (6 passed), new module tests/test_nonnormality_oracle.py (26 passed).

• Independent-oracle cross-checks for the spectral diagnostic layer (tests/test_qutip_spectral_oracle.py). The previous QuTiP cross-checks only validated the Liouvillian builder (matrix construction); the diagnostic outputs (D1 gap, D3 oscillating gap, steady state, GNS/KMS symmetrised gap) were only asserted for self-consistency and loose magnitudes (abs(gap - 0.2) < 0.05 or abs(gap - 0.4) < 0.05). The new module pins them on three canonical GKSL systems (amplitude damping, coherently driven dephasing, detailed-balance thermal qubit) against two independent oracles: closed-form analytic spectra (always runs) and qutip.liouvillian(...).eigenenergies() / qutip.steadystate(...) (@qutip_required). Includes the Mori-Shirai prediction that the symmetrised gap coincides with the standard gap at equilibrium (PRL 130, 230404 / arXiv:2212.06317). Test-only; no production code, anchor, or DiagnosticReport output changes. Verified locally via the CI command chain on CPython 3.14.4: ruff check src tests (exit 0), mypy src/liouscope (exit 0), anchors (21 passed), full suite with coverage (281 passed, 1 skipped — pre-existing local jsonschema-extra skip, coverage 93.54% ≥ 80% gate), pytest -m qutip (6 passed).

• CI test matrix extended to Python 3.14 (.github/workflows/ci.yml) and 3.14 added to the pyproject.toml Trove classifiers. Verified locally via the exact CI command chain on CPython 3.14.4: pip install -e .[dev,qutip] (qutip 5.3.0 ships cp314 wheels), ruff check src tests benchmarks (exit 0), mypy src/liouscope (exit 0), anchor regressions (21 passed), full suite (274 passed, 1 skipped, coverage 93.54% ≥ 80% gate). Closes the freshness gap where the library ran clean on 3.14 but CI never exercised it.

Documentation

• Added docs/RELEASE_AUDIT_v0.4.1.md, the post-v0.4.1-tag public/citable-release readiness audit (archive/provenance gates + post-v0.4.1 CI hardening). It is documentation-only and is not part of the tagged v0.4.1 source snapshot (the v0.4.1 tag at commit 1965f2b predates this file); it lives in [Unreleased] as accompanying post-release documentation. Updated docs/CANON_STATUS.md §5 to point public/citable-release work at this v0.4.1 audit instead of the superseded docs/RELEASE_AUDIT_v0.4.0.md §5.

Fixed

• codemeta.json was stale: it reported version: "0.2.0", described only "twenty diagnostics D1-D20", and set codeRepository to the non-existent github.com/coworker-research/liouscope. Synchronized to the repo canon: version → 0.4.1 (matching src/liouscope/_version.py and CITATION.cff), diagnostic description → 24 diagnostics D1-D24 (D1-D20 original submission set, D21-D24 post-submission; schema D1-D24-Übersicht-v3, per _consts.py / MANIFEST_SCHEMA.json), and codeRepository → github.com/marcohost33-maker/Liouscope. Metadata-only; no runtime/API change.
• pyproject.toml [project.urls] (Homepage/Repository/Issues) pointed to github.com/coworker-research/liouscope, which does not exist (HTTP 404 — the coworker-research org has no such repo). The published wheel/sdist metadata therefore carried dead project links. Corrected to the canonical repository github.com/marcohost33-maker/Liouscope, matching CITATION.cff, AGENTS.md ("Visibility: PRIVATE (marcohost33-maker/Liouscope)"), and the CHANGELOG.md version-compare links. Metadata-only; no runtime/API change.

0.4.1 — 2026-06-16

Release cut: numerics correctness + production-hardening that shipped on main after v0.4.0, PRs #43–#45 (resolvent conjugate-transpose fix, resolvent hardening + large-matrix scaling, classifier taxonomy doc fix, and coverage lifts across the fit, Liouvillian, sparse, classification, and numerics layers). PATCH bump per SemVer: the change set is a numerics correctness fix plus tests and documentation corrections. The one new public symbol (numerics.resolvent.ResolventConvergenceWarning) is a diagnostic warning on a numerics utility, not a new feature on the diagnose() / DiagnosticReport API surface, and no anchor or report output changes — see the per-entry "anchors unaffected" notes below.

Fixed

• numerics.resolvent.resolvent_norm large-matrix branch (n > 128): the SuperLU power-iteration computed the wrong conjugate-transpose for the resolvent. lu.solve(y.conj()).conj() evaluates conj(A)^{-1} y, which equals the required (A^H)^{-1} y only for symmetric A; on the non-normal Liouvillians this kernel targets the returned ||(zI - L)^{-1}||_2 was wrong by ~50% (e.g. 0.550 vs the dense reference 1.181 on a random non-symmetric matrix). Fixed to solve the LU's conjugate-transpose system directly via lu.solve(y, trans="H"); the power-iteration estimate now matches a dense SVD reference to < 1e-6 relative across seeds. The function is a public numerics utility and is not on the diagnose() report path, so no anchor or DiagnosticReport output changes (the small-matrix dense branch, used for Hilbert dimensions up to 128, was already correct). (The power-iteration convergence handling is hardened further under "Changed" below.)

Added

• Regression tests for the fit-model layer (tests/test_models.py): the closed-form M0–M3b evaluations and all initial-guess seeds, including the log-linear M0 regression with its too-few-positive-samples fallback and the FFT-based M3b dominant-frequency pick (supplied-omega and short-signal branches). Raises fitting/models.py coverage from 69% to 97% (the only remaining line is a defensive else unreachable for equal-length inputs).
• Branch coverage for the foundational Liouvillian builder and steady-state solver (tests/test_lindblad.py): the jump_ops=None default, the order != "F" guard, zero-rate-jump skipping, the rate-length guard (distinct from the jump-shape guard), the non-square-superoperator rejection, the no-exact-null-space smallest-singular-vector fallback, and the traceless-null-space "cannot normalise" RuntimeError. Raises core/lindblad.py coverage from 81% to 99%.
• Fallback-path tests for the Prony M3b seed (tests/test_fitting.py): short-signal, non-uniform-sampling, and too-few-samples-for-model-order fallbacks — pinning the robustness guards that let prony_seed degrade gracefully instead of raising. Raises fitting/prony.py coverage to 88%.
• Edge-branch tests for sparse.build.build_sparse_liouvillian (tests/test_sparse.py): the order/shape/rate-length validation errors, the jump_ops=None and rates=None defaults, and zero-rate-jump skipping. Raises sparse/build.py coverage from 71% to 100%.
• Regression tests pinning the A1–A12 mechanism classifier decision tree (tests/test_classification.py): synthetic-input coverage of every _pick_a_class branch, the _pick_verdict_tier thresholds (including the EXCLUDED and UNDEFINED paths unreachable through natural confidence values), and the _confidence scoring rules. Raises classification.py coverage from 71% to 100%.
• Tests for the previously-untested large-matrix branches of numerics.resolvent (tests/test_numerics.py): the SuperLU sparse solve (n > 256) and the power-iteration resolvent norm (n > 128, the regression guard for the fix above), plus a clustered-singular-value case proving the norm stays accurate to < 1e-3 when the top two singular values nearly coincide.

Changed

• diagnostics.classification documentation corrected to match the authoritative _consts taxonomy. The module docstring previously described an unrelated "evidence families" scheme (F1=spectral, F4=resolvent, …) that contradicted _consts.F_FAMILY_DESCRIPTIONS, where F1–F5 denote the literature-anchored gap-failure mechanisms (F1 Mori-Shirai overlap PRL 125 230604; F2 skin effect PRL 127 070402; F3 symmetrised gap PRL 130 230404; F4 quantum Mpemba PRL 127 060401; F5 phantom relaxation arXiv:2306.07876 — all references web-validated). The misleading inline comment on the A3 branch ("F4 resolvent-amplified non-normality") was fixed: A3 = overlap/eigenvector- amplified (Mori-Shirai 2020) correctly maps to family F1, which is what the code already returned — a comment/doc defect, not a behaviour change. Added a guard test (test_family_citations_consistent_between_docstring_and_consts) pinning the docstring and _consts to the same citations so they cannot drift apart again. Also removed a dead f_family parameter from the private _pick_verdict_tier helper (it never influenced the verdict/tier). No classifier output changes; anchors unaffected.
• D11b/D12 resolvent diagnostics now scale to large Liouvillians. The inline per-frequency dense inverse + SVD in diagnostics.resolvent.resolvent_peak_curve (201 dense O(n^3) solves, intractable for larger lattices and a duplicate of the numerics utility) is replaced by a delegation to numerics.resolvent.resolvent_norm. For n <= 128 — which includes every anchor/example system — this is bit-identical (same dense inverse + SVD); for n > 128 it uses the SuperLU shift-and-invert power iteration, so the resolvent peak (D11b) and ridge FWHM (D12) become computable where the dense path previously could not finish. The peak matches the dense reference to machine precision; low-norm tail points of the profile inherit the power iteration's documented ~1e-3 lower-bound behaviour in the clustered regime.
• numerics.resolvent.resolvent_norm hardened for production: the power iteration now emits a ResolventConvergenceWarning (new, exported from numerics.resolvent) when it exhausts its budget on tightly-clustered top singular values (the returned value is then a documented lower bound), the iteration budget was raised 80 → 200 (moderately clustered spectra now converge; tight clusters improve from ~2e-4 to ~1e-6 relative error), and the size cutoffs / iteration constants are now named module constants instead of magic numbers. New tests cover the warning path and a convergent clustered case.
• numerics.resolvent.resolvent_norm docstring now documents the method as the standard pseudospectra shift-and-invert approach (Trefethen, Pseudospectra of Linear Operators, SIAM Rev. 1997) and records why plain power iteration is sufficient for the dominant value (value-convergence is robust to top singular-value clustering, unlike eigenvector-convergence — verified empirically), so Lanczos/svds is intentionally not used.
• Packaging metadata modernised to PEP 639: pyproject.toml now declares the license as the SPDX expression license = "Apache-2.0" with license-files = ["LICENSE"], and the deprecated License :: trove classifier is removed (the SPDX expression is the single source). The setuptools build requirement is raised to >=77.0 (first release with PEP 639 support). This clears the deprecation that setuptools enforces after 2026-02-18 for the old license = {text = ...} table, and is the same Metadata-2.4 machinery behind the twine check license-file note recorded for this release. No runtime, API, or dependency change.

0.4.0 — 2026-06-07

Release cut: everything below shipped on main between v0.3.0 (2026-05-28) and this tag (18 commits, PRs #34–#41 — audit waves 2026-06-04/06-06 + D14 physics-scaling). MINOR bump per SemVer: backward-compatible API additions (gap= forwarding, TransientGridWarning, SOURCE_DATE_EPOCH).

Fixed

• D14 transient time-grid physics-scaling (audit 2026-06-06, P2-1): diagnostics.transient.trans_amplitude_ratio (sup_t ||e^{tL}||_2) previously used a fixed coarse grid linspace(0.01, 5.0, 30). For systems whose relaxation timescale 1/Delta falls outside [0.01, 5] — i.e. the small-gap, strongly non-normal regime the diagnostic is meant to detect — the supremum was silently underestimated (measured 27.8% too low on an amplitude-damping channel with gap 0.01: fixed grid 1.0209 vs the true sqrt(2) = 1.4142, recovered by a dense oracle). The grid is now physics-scaled to the spectral gap (D1) and the numerical abscissa: a two-scale window [0, ~8/Delta] with a log-spaced early segment (to resolve sharp growth peaks) plus a linear late segment. Across damping rates spanning three orders of magnitude the auto-scaled D14 now matches a dense reference to < 1e-3 relative. An explicit t_grid= still overrides the scaling (backward compatible); when no gap and no grid are given, the legacy coarse grid is kept as a fallback. A TransientGridWarning is emitted when the propagator norm is still rising non-negligibly at the right grid edge (the returned sup is then a lower bound). compute_transient_layer forwards the gap so diagnose() benefits automatically. New tests in tests/test_transient.py pin the behaviour against a dense brute-force oracle; the QuTiP physics-kernel parity cross-checks remain green (the change does not touch the Lindblad builder, steady state, or spectrum).
• Version single-source (audit 2026-06-06, P0): pyproject.toml no longer carries a second hard-coded version literal. [project] now declares dynamic = ["version"] and [tool.setuptools.dynamic] reads the version from src/liouscope/_version.py (the documented single source). Previously pyproject.toml said 0.3.0 while _version.py (= runtime liouscope.__version__ and every manifest's framework_version) said 0.2.0, so a built wheel reported the wrong version and every run manifest recorded a wrong framework_version — a provenance bug for a tool that claims paper-grade reproducibility. _version.py is bumped to 0.3.0; a new test (test_version_single_source) asserts importlib.metadata.version("liouscope") == liouscope.__version__.
• Reproducibility claim precision (audit 2026-06-06, P1): the README claim that two runs produce "byte-identical manifests" was empirically false — the embedded wall-clock timestamp varies, so the manifest SHA differed each run. README now states the true property (byte-identical except for the recorded timestamp; run_id/input_hash are run-invariant) and both properties are gated by tests.
• Diagnostic-count consistency (audit 2026-06-06, P1): README headline said "twenty diagnostics D1-D20" while its own table and the code constant DIAGNOSTIC_SCHEMA_VERSION = "D1-D24-Übersicht-v3" run to D24. README and AGENTS.md now consistently say "24 diagnostics D1-D24 (D1-D20 submission set; D21-D24 post-submission)".

Added

• SOURCE_DATE_EPOCH support in the manifest writer (audit 2026-06-06, P1): when this reproducible-builds standard env var is set to a fixed Unix timestamp, the manifest's timestamp field uses it instead of the wall clock, making the manifest fully byte-identical across runs. An unparseable/negative value is rejected fail-closed (no silent fallback to the wall clock). Gated by test_manifests_byte_identical_with_source_date_epoch.
• Boundary-guard hardening wave (audit 2026-06-06): fail-closed input validation and structured IO errors at the public surface. No API breaks, no new mandatory dependencies, anchor regressions unchanged.
  • numerics.linalg.require_finite_square_2d: a reusable boundary validator that rejects non-finite (NaN/inf), non-square, or empty operators with a structured, argument-named ValueError (e.g. "L_super contains non-finite entries (1 NaN, 0 inf)"). Previously such inputs flowed into scipy.linalg.expm/svd and surfaced as opaque, location-blind LAPACK messages ("array must not contain infs or NaNs" / "SVD did not converge") that named neither the offending argument nor the real defect.
  • diagnose() now validates L_super and any caller-supplied rho_initial / rho_steady_state (finiteness + shape match against d) at the entry boundary, before any numerics run. The previously uncaught path was supplying rho_steady_state (which bypasses the steady-state SVD), where an inf/NaN in L_super was only caught deep inside expm.
  • io.export.load_report fails closed with structured, path-bearing errors: a missing path raises FileNotFoundError("report file not found: ..."); malformed JSON or a non-object top level raises ValueError naming the file. Replaces the raw json.loads(Path(path).read_text(...)) that gave callers no context (the "exists()-then-read-raw" fail-open class).
  • io.export.dump_report and io.manifest.dump_manifest now create missing parent directories (parents=True, exist_ok=True) so nested artefact paths do not fail with a bare FileNotFoundError on first write.
  • tests/test_input_guards.py (12 tests) and tests/test_export.py (8 tests): negative/edge inputs (NaN, inf, empty, wrong shape, missing file, malformed/non-object JSON) fail before the happy path; valid input is not rejected (no false positives). Closes the prior coverage gap for io.export.
• Statistics-hardening wave (audit 2026-06-04, findings S1-S6):
  • core.lindblad.DegenerateSteadyStateError + steady_state(allow_degenerate=) (S1): a multi-dimensional Liouvillian null space (non-unique NESS / decoherence-free subspace) now fails closed instead of silently returning an arbitrary representative; opt in via allow_degenerate=True for one representative with a RuntimeWarning.
  • fitting.neff.ar1_correlation_corrected (S2): small-sample bias-corrected lag-1 autocorrelation (rho*(n-1)+1)/(n-3) (Marriott-Pope/Kendall; arXiv:2010.05870) + RuntimeWarning at n <= 40. Wired into fit_gls_ar1 so AR(1)-whitened CIs are no longer over-confident at small n.
  • Two exact analytic anchors in tests/test_anchors.py (anti-circularity): pure-dephasing Liouvillian gap = 2*gamma, amplitude-damping coherence decay rate = gamma/2, both asserted to atol=1e-12.
  • .github/workflows/ci-qutip.yml: dedicated job that installs the QuTiP extra and runs the QuTiP cross-checks without skipping (pytest -m qutip), plus a guard against a vacuous 0-collected run. Closes the gap where cross-family validation only ever skipped in CI. (Required-check registration is done separately by a maintainer.)
  • _zhou.CLAIM_STATUS / _zhou.CLAIM_REFERENCE (S6): the D24 Zhou predictor is marked pending/unverified because its cited reference (arXiv:2601.06256) could not be independently verified at audit time. README + module docstring annotated accordingly.
• _consts.EPS_DIV: canonical division-by-zero floor (1.0e-300) shared by the Petermann inner-product guard in diagnostics.nonnormality.petermann_factors and _zhou. Replaces the previously hard-coded 1.0e-300 magic numbers. Deliberately distinct from the physics-scale EPS_GAP/EPS_SUPP.
• tests/test_zhou.py: a closed-form anchor for the D24 Zhou predictor (single-qubit pure dephasing, gap = 1, K = 1 -> both bounds = log(1/eps)) plus a defective-mode guard test (a near-defective mode must not poison the finite upper bound).
• fitting.prony._default_seed and guarded prony_seed: the Prony seed now catches LinAlgError/ValueError from lstsq/np.roots on near-singular Hankel data (e.g. all-NaN/inf signals), emits a RuntimeWarning, and falls back to a safe default seed with a strictly positive amplitude. New regression tests in tests/test_fitting.py (fails-before on the pre-guard code, which raised LinAlgError on non-finite input).

Changed

• fitting.bootstrap.bca_ci (S3, S4): the BCa bias-correction z0 now uses the Efron-1987 half-correction for ties (mean(<) + 0.5*mean(==)) instead of a strict < (which drove z0 -> -inf when bootstrap replicates equalled theta_hat); CI endpoints now use linear quantile interpolation (np.quantile) instead of granular nearest-rank.
• CI mypy src/liouscope is now an enforcing gate (continue-on-error removed); the previously type-blind 18 mypy findings were fixed (annotated numpy return locals, is_dataclass instance narrowing). No behaviour change.
• _zhou: zero-eigenvalue and division-by-zero thresholds now use the canonical EPS_GAP/EPS_DIV constants instead of inline 1.0e-10/1.0e-300.
• _zhou.CLAIM_STATUS (S6 re-audit 2026-06-04): pending/unverified -> reference-verified-bound-coarser. The cited reference was independently verified against the arXiv PDF: Yi-Neng Zhou, "Universal Predictors for Mixing Time more than Liouvillian Gap", arXiv:2601.06256 (v3 2026-05-20, University of Geneva). The placeholder title and UNVERIFIED marker are replaced with the real title/author/version. The implemented upper bound is in the same family as Zhou's central result Eq.(16) and exact in the normal-mode limit (pure-dephasing anchor), but is a related, generally coarser surrogate: it uses the Petermann (Schatten-2) factor sqrt(K) rather than Zhou's per-mode trace-norm factor C_j = ||rho_j||_1 * ||sigma_j||_op, a single global gap/K_max instead of a per-mode maximum of (1/lambda_j) log(N_mode C_j), and omits the N_mode factor. Differences documented exactly in the _zhou module docstring; CLAIM_REFERENCE, README, and the status-lock test updated accordingly. (No formula change.)
• fitting.neff.ar1_correlation_corrected docstring hardened (no formula change): added a validity note (first-order Kendall / Marriott-Pope correction; reliable up to rho ~ 0.85, residual bias beyond that set by the truncation order) and full references (Marriott & Pope 1954; Kendall 1954; arXiv:2010.05870; Dou et al. 2026, Br. J. Math. Stat. Psychol.). Documents that a higher-order Kendall variant tested marginally better at high rho/n but was deliberately not adopted to keep the audit formula bit-stable.

0.3.0 - 2026-05-28

Added

• liouscope._zhou: Zhou universal mixing-time predictor (D24) as an opt-in diagnostic. Frozen ZhouPredictorResult dataclass.

Changed

• MANIFEST_SCHEMA.json moved from the repo root into src/liouscope/MANIFEST_SCHEMA.json. pyproject.toml already listed it under [tool.setuptools.package-data], but the file was not actually at that path, so wheels built from PyPI shipped without it. The schema is now correctly bundled (verified by inspecting the wheel) and loaded via importlib.resources so the lookup works under editable, wheel and zipfile installs.
• validate_manifest now uses a cached jsonschema.Draft202012Validator (per the python-jsonschema performance guidance) instead of the autodetecting jsonschema.validate convenience wrapper. The bundled schema declares draft/2020-12, so this is the matching validator class.
• _utc_now_iso() uses the idiomatic isoformat(timespec="microseconds").replace("+00:00", "Z") pattern. Pinning timespec guarantees a fixed-width 27-character timestamp string on every call, eliminating the case where zero-microsecond timestamps would lose the fractional component.
• liouscope.io.manifest_payload: schema-compliant projection of a DiagnosticReport that includes schema_version, taxonomy_version, and diagnostic_schema_version as documented in MANIFEST_SCHEMA.json.
• liouscope.io.dump_manifest: writes the manifest payload to a JSON file.
• liouscope.io.validate_manifest: validates a manifest dict against MANIFEST_SCHEMA.json. Uses jsonschema when available, falls back to a built-in subset check otherwise.

Fixed

• liouscope.io.build_manifest now uses timezone-aware datetime.datetime.now(UTC). The previous datetime.utcnow() call was deprecated in Python 3.12 and slated for removal in 3.14.
• liouscope._zhou.mixing_time_upper_bound rescaling formula. The previous version contained a no-op (epsilon / epsilon) and dropped the 1/gap factor, returning incorrect mixing-time estimates for any eps other than the original one. ZhouPredictorResult now carries the spectral gap and Petermann factor used to build it so rescaling is well-defined.
• README quickstart code now uses the actual public API (one_d_chain, heisenberg_xxz_hamiltonian, boundary_dephasing_jumps, diagnose(L, rho_initial=...), report.relaxation.beta_D, report.relaxation.bca_ci_beta) — the previous snippet referenced symbols that did not exist (Chain1D, XXZ, tau_eff, ci95).

Security & CI Hardening (Welle 4)

• zizmor workflow security audit added (SHA-pinned zizmorcore/zizmor-action@v0.5.6 via 5f14fd08...).
• All Actions SHA-pinned (actions/checkout@v4.2.2, setup-python@v6.2.0, etc) — Welle G Gold-Standard pattern.
• Dependabot configured (weekly grouped pip + github-actions, 7-day cooldown against Shai-Hulud-style supply-chain attacks, May 2026).
• OpenSSF Scorecard workflow (private-repo SARIF guard + workflow_dispatch).
• Tier-2.5 Branch Protection: enforce_admins=true, required_conversation_resolution=true, dismiss_stale_reviews=true (solo-dev pattern, required_approving_review_count=0).
• delete_branch_on_merge=true.
• .gitattributes (eol=lf) for cross-platform consistency.

0.2.0 -- 2026-04-17

Added

• Twenty diagnostics D1-D20 organised in six layers S/N/R/U/C/G.
• Twelve-class mechanism taxonomy A1-A12 (TAXONOMY_VERSION = "A1-A12-v3.1").
• Fit hierarchy M0/M1/M2/M3a/M3b with Prony-seed initialisation for M3b.
• Statistical pipeline: GLS with AR(1) residuals, N_eff via Geyer 1992 IPS estimator, AICc with N_eff correction, parametric bootstrap with BCa confidence intervals.
• Sparse path (liouscope.sparse) with ARPACK shift-invert for d up to 128.
• Run manifest with SHA-256 run-id and JSON export. Schema version 1.2.0.
• Four lattice geometries (1D chain, 2D square, honeycomb, triangular) and four benchmark Hamiltonians (Ising, XY, Heisenberg-XXZ, Bose-Hubbard).
• Three dissipator families (bulk, boundary, engineered).
• Five validation systems V1-V5 as library functions in liouscope.examples.
• Paper figure pipeline.

Fixed

• FIX-1: GNS Gram matrix is rho_ss^T (x) I (not the KMS form).
• FIX-2: Column-stacking via flatten(order='F') is enforced everywhere.
• FIX-3: LEP detection includes complex-conjugate eigenvalue pairs.
• FIX-4: Pauli-sector rate is labelled distinctly from Delta_s.
• FIX-5: M0 fit uses log-linear regression on D(rho||pi) as baseline.
• FIX-6: Henrici eta_N via Schur decomposition.
• E1-E10: All ten normative patches from the v3 audit applied.

Correctness anchors enforced (regression tests in tests/test_anchors.py)

• A. Column-stacking order='F'.
• B. GNS Gram rho_ss^T (x) I.
• C. Alicki adjoint direction L_tilde* = (rho^{-1} (x) I) L (rho (x) I).
• D. zgeev (scipy.linalg.eig) for non-Hermitian Liouvillian.
• E. SuperLU for resolvent computations.
• F. AICc-only model comparison (M1/M2 are not nested).
• G. Parametric bootstrap on GLS-AR(1) residuals with BCa-CI.
• H. N_eff via Geyer 1992 IPS in AICc small-sample correction.
• I. Conjugate-pair inclusion in LEP proximity.
• J. supp(rho_0) subset supp(rho_ss) check with eps = 1e-12 regularisation.
• K. HS adjoint distinct from pi-weighted adjoint.
• L. TAXONOMY_VERSION stamped on every ClassificationResult.
• M. D11 = Bohr-AP (Basso 2025), D11b = resolvent peak.
• N. D24 = Zhou (universal mixing-time predictor), not Lee-Bound.