feat(#117,#168): conditioning audit with Schur-projector cluster semantics by marcohost33-maker · Pull Request #173 · marcohost33-maker/Liouscope

marcohost33-maker · 2026-09-27T00:19:17Z

Summary

Current-main consolidation of the audit-only zero-mode conditioning work from #166 plus the defective-cluster correction from #172.

This PR is intentionally rebuilt non-destructively from the current protected main after #171 merged. It preserves the #152 finite-generator guard and avoids rebasing/force-updating the older stacked branches.

Scientific contract

Simple zero mode

scalar reciprocal eigenvalue condition is the direct |y^H x|;
scalar first-order structural/solver displacement estimates remain available when finite;
the independent Schur path is not forced onto a simple root, avoiding unnecessary GEES/Sylvester round-off.

Multi-mode zero cluster

reorder the selected cluster in a complex Schur form with LAPACK xTRSEN;
use xTRSEN for reordering only;
solve T11 R - R T22 = T12;
first require the independent Schur spectrum to match the report spectrum mode-for-mode inside the existing round-off agreement band; a same-size but numerically different cluster fails closed;
report the reciprocal spectral-projector 2-norm 1/sqrt(1 + ||R||_2^2);
do not identify xTRSEN.S with that exact quantity: it is a Frobenius-norm lower bound;
return CLUSTER_ONLY and withhold scalar first-order eigenvalue-displacement claims, because defective versus semisimple multiplicity is not safely identifiable from numerical proximity alone.

Fail-closed rules

accepted/report spectrum must agree with the audit solve before scalar evidence is attached;
non-representable structural evidence returns UNAVAILABLE, including in cluster-only mode;
malformed/non-finite cutoffs are rejected;
no filter, D1-D24 value, certificate, classifier verdict or tier consumes this audit.

Regression discrimination

canonical fix(#122): repair the two-scale relaxation window instead of disclosing it #127 non-normal simple-root fixture retains its measured s≈2e-7 scalar path;
exact Cluster conditioning from a reordered Schur decomposition, not from eigenvectors #168 Jordan-cluster fixture is invariant under unitary re-expression and reports CLUSTER_ONLY;
constructed Schur fixture distinguishes the exact projector 2-norm from LAPACK's Frobenius lower bound;
semisimple repeated cluster cross-checks Schur-projector conditioning against the legacy invariant-subspace formula;
stiff Stiff generators: dense eigensolver displaces the whole slow spectrum beyond any threshold's reach — structural deflation needed #112 deflation failure remains a negative control: conditioning does not claim to detect it;
audit on/off leaves reported diagnostics and certificate bit-identical.

Provenance / current-main consolidation

Source work:

feat(#117): audit-only eigenvalue-conditioning evidence for the zero mode #166: issue Zero-mode conditioning: is the structural certificate enough for stiff non-normal generators? #117 audit-only conditioning instrument and nine rounds of review hardening;
fix(#168): Schur-projector conditioning for repeated zero clusters #172: issue Cluster conditioning from a reordered Schur decomposition, not from eigenvectors #168 Schur-projector correction and cluster-only claim boundary;
fix(#152): fail closed on non-finite Liouvillian assembly from current main #171 is already merged into main and is retained by construction.

CHANGELOG.md and CITATION.cff were merged onto current-main content rather than copied wholesale from the stale stack; comparison against main shows additions only in both files.

Schema / release impact

Methodology and audit-field semantics change, so README / CHANGELOG / CITATION are updated. The run-manifest contract is unchanged and MANIFEST_SCHEMA is deliberately not bumped.

Test plan

Supersedes #166 and #172 for integration once this current-main PR is green.

Refs #112, #117, #127, #130, #168. The default call is now basis invariant (ABSTAIN without a cutoff); the remaining finite-cutoff limit is tracked in #183, so #168 stays open until that is closed.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 41240d54ee

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Both sides add independent Unreleased entries (#117/#168 vs #162/#177); kept both, no text changed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Without zero_tolerance the audit used an internal NaN cutoff. Every ordered comparison with NaN is False, so the scalar path could only answer BENIGN (diag(0,-1,-1,-2), trace defect 1.414, came back BENIGN), and the zero set was chosen by proximity, which a defective double root split by ~sqrt(eps) makes basis dependent (issue-#168 pair: CLUSTER_ONLY plain, BENIGN rotated; Codex review thread on conditioning.py:809). - default call returns verdict ABSTAIN / reason zero_tolerance_not_supplied: observations only (smallest eigenvalue, magnitude, trace defect), every zero-set quantity NaN, never BENIGN - scalar verdict via _scalar_verdict: BENIGN only on a TRUE budget <= tol; NaN/inf budget -> CONDITIONING_LIMITED, non-finite cutoff -> ABSTAIN - tests: default invariance on the #168 pair, non-trace-preserving default, 12-operator negative sweep, verdict truth table; measurement tests now pass the layer's own cutoff; longdouble fixtures skip where longdouble==float64 - benchmark families B/D/G follow the new default Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

README, CHANGELOG (Unreleased/Fixed) and CITATION note: a call without zero_tolerance returns ABSTAIN and never BENIGN; the spectral layer always passes its certified cutoff, so no reported number moves. Full-suite evidence run follows in chunks (~15 min). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

chatgpt-codex-connector · 2026-09-30T09:30:03Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

…the #183 limit Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

chatgpt-codex-connector · 2026-09-30T09:45:48Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

…ling Clean merge; ruff, mypy (58 files) and the touched test files (gls/neff + issue117/issue168 conditioning) green locally. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

marcohost33-maker added 22 commits September 27, 2026 02:17

port(#117,#168): README.md

376dd22

port(#117,#168): CITATION.cff

a001ea7

port(#117,#168): benchmarks/issue117_zero_mode_conditioning.py

a00d5f0

port(#117,#168): src/liouscope/_consts.py

f60c305

port(#117,#168): src/liouscope/_types.py

df81b4c

port(#117,#168): src/liouscope/diagnostics/spectral.py

ff66ecc

port(#117,#168): src/liouscope/numerics/__init__.py

586f76c

port(#117,#168): src/liouscope/numerics/conditioning.py

daccac1

port(#117,#168): tests/test_issue117_conditioning.py

d938df1

port(#117,#168): tests/test_issue168_schur_conditioning.py

e91a060

docs(#117,#168): merge audit citation onto current main

89ffaef

docs(#117,#168): merge audit changelog onto current main

6be6540

fix(#168): bind Schur projector to the reported spectrum

f2fa0f1

test(#168): bind focused Schur checks to reference spectra

f091927

test(#168): fail closed when Schur and report spectra diverge

2986b72

docs(#168): document Schur-to-report spectrum agreement gate

df6ec82

docs(#168): record Schur spectrum agreement fail-closed rule

fc747c9

docs(#168): bind projector claim to spectrum agreement

46fb950

docs(#168): distinguish simple scalar and cluster Schur paths

48fb3d4

test(#168): assert no scalar claim across Schur availability

6436b4e

test(#168): reject same-size projector from a different spectrum

0ee425f

fix(#168): reject unavailable structural evidence before Schur audit

41240d5

marcohost33-maker marked this pull request as ready for review September 27, 2026 00:48

chatgpt-codex-connector Bot reviewed Sep 27, 2026

View reviewed changes

Comment thread src/liouscope/numerics/conditioning.py

marcohost33-maker and others added 3 commits September 30, 2026 10:56

Merge origin/main into #173 (resolve CHANGELOG/CITATION as union)

39d6a76

Both sides add independent Unreleased entries (#117/#168 vs #162/#177); kept both, no text changed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

marcohost33-maker mentioned this pull request Sep 30, 2026

Conditioning audit: finite cutoff below the defective split is still basis dependent (#168 follow-up) #183

Open

docs(changelog): mark the #173 default-call change as breaking; note …

083da7d

…the #183 limit Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

marcohost33-maker merged commit 39682ad into main Sep 30, 2026
34 checks passed

marcohost33-maker deleted the codex/issue117-168-current-main-20260927 branch September 30, 2026 09:52

marcohost33-maker mentioned this pull request Sep 30, 2026

feat(#117): audit-only eigenvalue-conditioning evidence for the zero mode #166

Closed

14 of 19 tasks

Navigation Menu

Repository navigation

feat(#117,#168): conditioning audit with Schur-projector cluster semantics - #173

feat(#117,#168): conditioning audit with Schur-projector cluster semantics#173
marcohost33-maker merged 26 commits into
mainfrom
codex/issue117-168-current-main-20260927

marcohost33-maker commented Sep 27, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

chatgpt-codex-connector Bot commented Sep 30, 2026

Uh oh!

chatgpt-codex-connector Bot commented Sep 30, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

marcohost33-maker commented Sep 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Scientific contract

Simple zero mode

Multi-mode zero cluster

Fail-closed rules

Regression discrimination

Provenance / current-main consolidation

Schema / release impact

Test plan

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot commented Sep 30, 2026

Uh oh!

chatgpt-codex-connector Bot commented Sep 30, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

marcohost33-maker commented Sep 27, 2026 •

edited

Loading