Repository navigation
feat: anisotropic Čech filtration value and gradient for ellipsoid simplices (ported from ellcech) - #140
feat: anisotropic Čech filtration value and gradient for ellipsoid simplices (ported from ellcech)#140uda-lab-agent wants to merge 29 commits into
Conversation
…(ported from ellcech) Squash of local topic branch project26-port-minimax (8 commits) after two local Codex review rounds (gpt-5.6-sol, read-only): - round 1: REQUEST CHANGES, 3 documentation findings (contract exactness under regularization/conditioning; regularization scope and condition_number_limit guarantee; nonexistent damped-Newton fallback); fixed in 972ddbc (gpt-5.6-luna). - round 2: REQUEST CHANGES, residual wording (unadjusted := regularization == 0 and condition_number_limit is None; outputs are approximations); fixed in 2ed9783. Two-round cap reached; dispositions recorded here. No executable change relative to ellcech 82d13e3 (650-case bitwise comparison); checklist black/flake8/mypy/stubtest/pytest(396) passes. Names are provisional pending owner decisions in uda-lab/project-ellphi#26. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
…headers Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
…ngency_simplex_grad) Squash of local topic branch project26-pairwise-contract (6 commits) after two local Codex review rounds (gpt-5.6-sol, read-only): - round 1: REQUEST CHANGES — blocker: engine ignores the packed constant term, so t/point/dt_dcoef are valid only for normalised rows; major: SimplexTangencyGrad field order; minor: non-finite tolerances; minor: right-triangle fixture assertions. Fixed in 81b311c (gpt-5.6-luna): per-row normalisation gate (ValueError otherwise), documented normalisation-preserving derivative, field order, tolerance validation. - round 2: REQUEST CHANGES — blocker: catastrophic cancellation when evaluating the packed quadric at far-translated points (t from result.alpha, centred constraint evaluation); major: gate rejected valid coef_from_cov output for ill-conditioned covariances and d=1. Fixed in 9cca0ca: conditioning-aware bound max(norm_tol*max(1,|c|), 64*eps*cond2(A)*max(1,|c|,|b^T A^-1 b|)); d=1 packed conics are a pre-existing EllPHi limitation (unpack_conic), documented and reported with a clear error. Two-round cap reached; dispositions recorded here; checklist re-run independently (446 passed). Public contract: SimplexTangencyResult(t, point, mu, support, active_set) with active_set = tight constraints (paper sense) and support = weight support; raises on non-convergence; SimplexTangencyGrad(t, point, mu, dt_dcoef, support, active_set) in coef space. Engine moved to private _minimax_python / _minimax_grad_python modules (history preserved). General constants inside the solve are implemented on the engine branch and will lift the normalisation gate. Names provisional (#26). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
…sets) and add the public surrogate fixture Squash of local topic branch project26-engine-hardening (7 commits) after two local Codex review rounds (gpt-5.6-sol, read-only): - round 1: REQUEST CHANGES — major: snapping roundoff-sized constants falsified the packed-coordinate gradient; major: negative engine value clipped to t=0 with inconsistent sets; minor: offsets not checked finite before the k==1 fast path; minor: incomplete deviation list; nits: empty face misreported, fixture provenance not reproducible. Fixed in e8a2660 (gpt-5.6-luna): exact offsets passed through, ValueError for negative values beyond roundoff, finite check, deviation list, empty_face status, generation recipe + regeneration test. - round 2: REQUEST CHANGES — major: negative tolerance rejected legitimate cancellation-scale values (centre (1e6,-1e6)); minor: post-polish Cholesky failure deviation missing. Fixed in cc61adc: cancellation bound value_roundoff = sum_i 64*eps*cond2(A_i)*max(|c_i|, |xbar_i^T A_i xbar_i|) carried from _prepare_coefs; deviation list completed; the 1e8 translated test rewritten to a well-scaled claim (packed constants carry roundoff of order eps*|xbar|^2; a representation limit, not a solver property). Two-round cap reached; dispositions recorded here; checklist re-run independently (481 passed). Engine deviations from ellcech 82d13e3 are enumerated in the module docstring: Armijo-rejected steps never applied (best iterate kept, converged=False), face re-identification after accepted steps, undamped Newton stops on ill-conditioned reduced Hessian, newton-cold falls back to a fresh FW iterate (converged=False), failure statuses force converged=False, post-polish Cholesky failure forces converged=False, full-step Armijo roundoff allowance, additive offsets (delta_i) with d alpha/d delta_i = mu_i. Fidelity: 3,048 legacy converged=True results over 480 instances identical to 1e-12 (scripts/compare_minimax_engine.py, opt-in). Public surrogate fixture (k=6, d=2, deterministic recipe) encodes the #27 mechanism: SLSQP meets value/gap bounds, every other method raises. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Codecov Report❌ Patch coverage is
📢 Thoughts on this report? Let us know! |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 50120234a7
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".
Validate solver controls and preserve a non-empty Frank-Wolfe face. Use the regularized dual for positive-regularization SLSQP and cover all three review regressions. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: dc39c2eb9e
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".
…hResult, cech_grad) Rename the public many-body API after the anisotropic Čech filtration. For k bodies, t^2 = alpha(sigma) is the least scale at which the growing ellipsoids share the critical point x*, where the active normals are positively dependent; the boundaries are not mutually tangent except in the pairwise k = 2 case. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 565547a112
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".
Thread matrix stabilization through Newton polishing and convergence checks, keep stabilized SLSQP objectives consistent, report successful FW fallbacks, and prevent thresholded-face false convergence. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
Remove stabilization keywords from the public value and gradient APIs, update their stubs and guide, and cover rejected controls plus the public fallback path. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c878998467
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".
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6d0e6fe361
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".
…ro conditioning escalations Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9c17323384
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".
…lised Newton residuals Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ce78b372dc
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".
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
|
Correction to the evidence comment above: review #140 (review) on |
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
|
Codex Review: Something went wrong. Try again later by commenting “@codex review”. ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
… alpha Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
|
Codex Review: Something went wrong. Try again later by commenting “@codex review”. ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3bf2406f38
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".
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
|
Codex Review: Something went wrong. Try again later by commenting “@codex review”. ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a401b9c0c7
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".
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e5a5e9148b
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".
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
|
@codex please review this PR |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 99a0dfa461
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".
|
Updated the PR body with the settled named-solver failure contract, transparent composite/default fallback policy and ordinary-input performance priority; value/status fidelity remains separate from derivative guarantees, and no solver choice was imposed. |
…eclare line-search safeguard Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
…tics Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM
… fields and the auto route Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4o
|
@codex please review this PR |
|
Codex Review: Didn't find any major issues. Nice work! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
|
Reviewed-by: chatgpt-codex-connector[bot] — review #140 (comment) on db5ea80: 0 inline comments. |
|
Added an open acceptance item to the PR body: at db5ea80, explicit scipy-slsqp can invoke Newton without exposing that transition in public diagnostics; named-method identity and traceability still need to satisfy the settled contract. |
Governing issue: uda-lab/project-ellphi#26 (private). Do not merge: the owner will decide the target release branch and give explicit merge approval; until then this PR stays open for review. Based on ellphi-0.1.3 only because that is the current development branch; the base will be retargeted when the owner names the release.
What this adds
Settled owner numerical contract (2026-10-05)
This is the acceptance direction for this PR; implementation compliance still requires commit-scoped review and tests.
Review/test ordinary-input speed and achieved accuracy; meaningful failure behavior and named-method fidelity; and, separately, any composite/default fallback triggers, route visibility, final checks and total cost. Retain bounded boundary/conditioning/degeneracy tests with explicit expected success or failure behavior. An honest exception can satisfy a failure-behavior test but is still a failure in numerical-success reporting.
Keep the narrowed value/status fidelity gate separate from derivative guarantees. A success flag or small gap does not certify derivative accuracy; retain separate derivative evidence and the documented support, locality, nondegeneracy and linear-solve hypotheses.
FW remains a provisional candidate, not an owner preference or mandated default, warm-up or fallback. This clarification selects no default, replacement method, tolerance or acceptance-rate target. Earlier review replies saying “no automatic fallback” must be read under this layer-specific contract: hidden substitution inside a named solver is prohibited; transparent, reasonable fallback at a composite/default layer is allowed. Historical candidate/default reports do not impose a solver choice.
The owner intends this responsibility split to match EllPHi's standard. This body update does not verify the existing implementation, change code or tests, launch experiments/review dispatch, or expand merge/release authorization.
Open named-method acceptance item (source check: 2026-10-05)
scipy-slsqphonor the named-method contract. Atdb5ea80ec3ee87e4bfa20bf9ba4adc46be9ade7a, the engine's SLSQP branch calls_newton_polishwhen SLSQP has not met its success/gap criteria and recordsaccuracy_polish="newton"internally. The public stage construction and returnedCechInfoomit that transition and identify the method asscipy-slsqp. The public interface supplies no disable switch for this polish; validation requiresnewton_max_iter > 0. Thus strict identity/traceability is not yet established for every named method. Explicitscipy-slsqpmust either succeed within its declared method or raise a meaningful exception. Any retained SLSQP+Newton composite must be explicitly identified and externally traceable under the policy above. Verify the resulting route identity, transition diagnostics and failure behavior with focused tests; this does not require every boundary case to return a value. The existence ofautoor green CI does not close this specific acceptance item. This is a source-level finding, not a new runtime experiment or an implementation prescription.Review history
Three local pseudo-PR rounds with Codex (gpt-5.6-sol) reviews, two rounds each; findings and dispositions are recorded in the squash commit messages.
Checklist
black, black --check, flake8, mypy (src+tests), stubtest with the existing allowlist, pytest 497 passed, uv lock --check. All required local checks passed.
Closes nothing; tracked in uda-lab/project-ellphi#26.
🤖 Generated with Claude Code
https://claude.ai/code/session_01Mn94Zd2yUjWGNQX3rtd4oM