docs: rebuild development roadmap for 0.2.4

[TheHiddenObserver]

Summary

Rebuild the development planning hierarchy from the validated statgpu 0.2.4 baseline and make the compact execution queue faithfully enforce the repository development gates.

This is a documentation-only planning change. It does not modify public APIs, numerical code, tests, packaging, or release artifacts.

Changes

• add dev/plans/README.md with authority separated by responsibility rather than one unsafe global precedence order;
• make applicable .claude workflow/skill and dev/AGENTS.md authoritative for hard development/completion gates;
• make maintained implementation/tests plus current user documentation authoritative for public capability;
• make ROADMAP.md authoritative only for priority and sequencing;
• make GitHub issues/PRs authoritative only for executable scope and dependencies;
• explicitly state that roadmap and issue scope may narrow work but may not weaken hard gates;
• add canonical dev/plans/ROADMAP.md with current priorities, dependencies, and definition of done;
• add dev/plans/ISSUES.md mapping roadmap work packages to executable issues and dependency order;
• replace the stale dev/plans/TO_DO.md with the mandatory gate checklist plus active queue;
• mark dev/plans/plan.md as a historical delta rather than a current priority source;
• classify older module plans as active design references, historical research plans, or archive material;
• base future work on the released and production-verified 0.2.4 commit 0aeeb95b60e3e274053b8f1b6427ae50c8eec015.

Multinomial staging decision

The roadmap now follows the non-tunable-first design:

• linear_model: design and implement unpenalized multinomial logistic regression phase 1 #96 is strictly unpenalized multinomial/softmax Phase 1 and exposes no L2 or other regularization parameter, so no multinomial CV surface is required in that issue;
• linear_model: implement complete penalized multinomial logistic suite with CV #98 is one complete penalized multinomial capability package, dependent on linear_model: design and implement unpenalized multinomial logistic regression phase 1 #96;
• linear_model: implement complete penalized multinomial logistic suite with CV #98 must cover at least L2, L1, ElasticNet, SCAD, and MCP across NumPy, CuPy, and Torch;
• every penalty exposed by direct fit in linear_model: implement complete penalized multinomial logistic suite with CV #98 must ship with path/grid, deterministic CV, selection, final refit, supported inference contract, external alignment, physical-GPU validation, and synchronized EN/CN documentation;
• linear_model: implement complete penalized multinomial logistic suite with CV #98 may use a bounded internal PR sequence, but it may not close or advertise completion after only L2 or only direct-fit support.

Executable issues

• benchmark: synchronize dashboard PR #76 with statgpu 0.2.4 master #90 — synchronize benchmark dashboard PR feat: benchmark dashboard frontend and canonical data pipeline #76 with current master;
• benchmark: add canonical cross-validation source and dashboard coverage #91 — add a canonical cross-validation benchmark source and dashboard coverage;
• benchmark: complete dashboard production QA and accessibility gate #92 — complete dashboard production QA and accessibility;
• panel: complete Tier-1 diagnostics and shared inference framework #93 — complete Panel Tier-1 shared framework, diagnostics, fit statistics, and covariance support;
• survival: implement Kaplan-Meier and Nelson-Aalen estimators #94 — implement Kaplan-Meier and Nelson-Aalen estimators;
• survival: implement initial parametric AFT model family #95 — implement the initial Weibull/log-normal/log-logistic AFT family;
• linear_model: design and implement unpenalized multinomial logistic regression phase 1 #96 — design and implement unpenalized multinomial logistic regression Phase 1;
• backends: define shared sparse-array contract without silent densification #97 — define the shared sparse-array/backend contract without silent densification;
• linear_model: implement complete penalized multinomial logistic suite with CV #98 — implement the complete penalized multinomial logistic suite with CV closure.

Validation

• branch agent/rebuild-roadmap-v0.2.4 is based directly on master commit 0aeeb95b60e3e274053b8f1b6427ae50c8eec015;
• the branch has one intentionally squashed documentation commit;
• changed paths are limited to:
  • dev/plans/README.md;
  • dev/plans/ROADMAP.md;
  • dev/plans/ISSUES.md;
  • dev/plans/TO_DO.md;
  • dev/plans/plan.md;
• no runtime behavior is changed;
• hosted Tests and Maintenance compatibility workflows are the authoritative merge gates and must pass on the final head.

Impact

After this PR, capability claims, development gates, priority, executable scope, and historical context each have a non-conflicting authority source. The roadmap no longer permits a tunable penalized direct-fit capability to be treated as complete while its CV path remains planned.

[TheHiddenObserver]

Executable roadmap issues created

The roadmap is now backed by scoped GitHub issues rather than an unstructured checklist:

• benchmark: synchronize dashboard PR #76 with statgpu 0.2.4 master #90 — synchronize benchmark dashboard PR feat: benchmark dashboard frontend and canonical data pipeline #76 with the 0.2.4 master baseline;
• benchmark: add canonical cross-validation source and dashboard coverage #91 — add a canonical cross-validation benchmark source and dashboard coverage;
• benchmark: complete dashboard production QA and accessibility gate #92 — complete dashboard production QA, cross-browser smoke, accessibility, and documentation integration;
• panel: complete Tier-1 diagnostics and shared inference framework #93 — complete Panel Tier-1 shared framework, diagnostics, fit statistics, and covariance support;
• survival: implement Kaplan-Meier and Nelson-Aalen estimators #94 — implement Kaplan-Meier and Nelson-Aalen estimators;
• survival: implement initial parametric AFT model family #95 — implement the initial Weibull/log-normal/log-logistic AFT family;
• linear_model: design and implement unpenalized multinomial logistic regression phase 1 #96 — design and implement multinomial logistic regression Phase 1;
• backends: define shared sparse-array contract without silent densification #97 — define the shared sparse-array/backend contract without silent densification.

dev/plans/ISSUES.md records the priority and dependency order. dev/plans/TO_DO.md now links each active work package to its issue.

[TheHiddenObserver]

Validation note

This PR is documentation-only. Runtime, package, and numerical test suites are not affected. The branch comparison against master is the required scope check; hosted documentation/static workflows remain the authoritative merge gate if triggered.

[TheHiddenObserver]

The issue index is intentionally limited to current executable work. Deferred module families remain in ROADMAP.md without open implementation issues until a concrete user need and scoped three-backend design exist.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[chatgpt-codex-connector]

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.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[TheHiddenObserver]

dev/plans/TO_DO.md has been expanded to reflect the blocking rules in dev/AGENTS.md and .claude/workflows/new-module-dev.md rather than only summarizing them.

The revised checklist now includes impact classification, the four hard exit statuses, three-backend and device-locality requirements, GPU memory lifecycle, architecture reuse, direct-fit/CV closure, strict inference, full formula semantics, objective and penalty-scale mapping, precision/convergence gates, architecture-specific tests, physical-GPU validation tiers, review/fix closure, benchmark provenance, EN-first/CN-follow documentation, changelog synchronization, and the required completion report.

The branch history was then squashed to one documentation commit based directly on the 0.2.4 master baseline so the PR contains no connector-generated empty commit noise.

[TheHiddenObserver]

Review result: not ready to merge

I found two blocking governance inconsistencies:

1. dev/plans/README.md currently says to use one global precedence order when documents disagree, but it places roadmap priorities and executable issues above dev/AGENTS.md / .claude development gates. That contradicts both the PR body and TO_DO.md, which correctly state that .claude hard protocols and dev/AGENTS.md take precedence over roadmap and issue scope. Please rewrite the README hierarchy as category-specific sources of truth, or make explicit that hard development gates override roadmap/issues whenever completion requirements conflict. An issue or roadmap item must never weaken a hard gate.

2. The multinomial roadmap and issue #96 expose L2-penalized direct fitting while explicitly deferring multinomial CV. L2 is a tunable public capability, so this conflicts with the direct-fit/CV closure rule in dev/AGENTS.md, .claude/workflows/new-module-dev.md, and the revised TO_DO.md. Resolve this before merge by choosing one compliant form: (a) Phase 1 is unpenalized only; (b) Phase 1 includes the L2 CV path; or (c) record an explicit user-approved CV deferral with user-visible behavior, tests, docs, and a follow-up issue, and make issue #96 non-closeable without that approval.

Everything else reviewed cleanly: the diff is documentation-only and scoped to five planning files; the branch contains one commit based on the 0.2.4 master baseline; issues #90–#97 exist; there are no unresolved inline threads; GitHub reports the PR mergeable; and final-head Tests #1412 plus Maintenance compatibility #425 both passed.

GitHub does not allow an author to submit REQUEST_CHANGES on their own PR, so this is recorded as a COMMENT review, but the two findings above are merge blockers.

[TheHiddenObserver]

Blocking review follow-up

Both previously identified governance blockers are resolved on head d0ee8f8cd1607eaf9d3cc2d57b80d27bd665e7ab.

1. dev/plans/README.md now assigns authority by responsibility. Applicable .claude workflow/skill and dev/AGENTS.md own hard completion gates; maintained implementation/tests plus user documentation own capability; ROADMAP.md owns priority; issues/PRs own executable scope. Roadmap and issue scope may narrow work but may not weaken hard gates.
2. Multinomial staging now follows the unpenalized-first option. linear_model: design and implement unpenalized multinomial logistic regression phase 1 #96 is non-tunable and strictly unpenalized. New issue linear_model: implement complete penalized multinomial logistic suite with CV #98 owns the complete penalized multinomial package, requires at least L2/L1/ElasticNet/SCAD/MCP, and requires direct-fit/CV/final-refit closure for every declared penalty before completion.

The branch remains a single documentation commit based directly on the 0.2.4 master baseline. Final hosted checks are running on this head.