[Process] Keep viable scientific PRs open and actively harvest or challenge canon

[vantasnerdan]

Positive objective

Make the scientific PR lifecycle advancement-oriented without weakening artifact correctness or claim promotion: a valuable PR with a finite repair, refactor, harvest, or governed canon-challenge path remains open; reviewers actively move correct reusable atoms toward main; and close-unmerged is reserved for explicit terminal dispositions.

Scope

• Correct the live PR Exact-mass one-loop proper-time coefficients with preregistered regulator comparison (P230, advances #76) #77 lifecycle error and preserve its valuable work.
• Revise AGENTS.md, AGENTS_START_HERE.md, research-pr-harvest, and physics-erdos-loop.
• Revise the continuation, claim-review, effort, campaign, and research-arc memory templates where they currently omit lifecycle and canon-challenge state.
• Add or extend process validation only where it proves the new invariant rather than counting prose.

Success gate

1. The process distinguishes request-changes, active refactor, active harvest, superseded-after-landing, terminal rejected/withdrawn, and merged states.
2. A reviewer may not close/delete a valuable source PR merely because it is not mergeable yet, lacks accepted dependency closure for promotion, or has a finite representation/provenance repair.
3. Every refactor then merge unit has an owner or explicit handoff, a live source/harvest PR, and a next executable action; the source stays open until value lands or terminal evidence is recorded.
4. Canon remains authoritative for current releases but explicitly revisable through challenges and separately governed foundational revisions; canon conflict blocks promotion, not truthful conditional artifact harvest or continued investigation.
5. Closure requires a unit-level terminal-close test and records what landed, what is irreducibly non-reusable, what superseded it, or who explicitly withdrew it.
6. Skills and memory templates validate, repository checks pass at the bounded documentation/process scope, and PR Exact-mass one-loop proper-time coefficients with preregistered regulator comparison (P230, advances #76) #77 is open with corrected coordination state.

Dependencies and authority

Depends on the existing three-way separation between artifact merge, claim promotion, and goal completion; the non-self-merge rule; PR #77 and goal #76 as the motivating case; and the current accepted release as authority for promotion only. No scientific claim or release changes in this issue.

Coordination boundary

Branch: process/pr-lifecycle-active-harvest. Intended write surfaces: AGENTS.md, AGENTS_START_HERE.md, .agents/skills/research-pr-harvest/SKILL.md, .agents/skills/physics-erdos-loop/SKILL.md, and relevant memory-templates/. Avoid concurrent edits to those files. The implementing agent will open an Advances PR and cannot merge it; a distinct reviewer/owner must merge.

[vantasnerdan]

Corrective process PR opened: #83

The final boundary changes the root contract, onboarding, PR template, research harvest and physics skills, five memory templates, and a contribution-policy regression test. It makes finite refactor/harvest work an active lifecycle state, adds a terminal-close test, keeps canon challengeable without weakening promotion authority, and scopes validation by measured impact.

Validation: scripts/validate.sh --pytest-scope tests/test_public_contribution_surfaces.py passed all fixed checks plus 9 targeted tests; both skill validators passed; GitNexus reports LOW risk and no affected process; final record-only updates passed memory and repository-schema validation; git diff --check passed separately. No full pytest replay was run or claimed.

Merge authority is handed to a distinct reviewer or repository owner.

[vantasnerdan]

Final CI audit found one remaining contradiction: .github/workflows/validate.yml still hard-coded --full for every PR. PR #83 now also replaces that with scripts/validate_changed.py, which conservatively chooses full, affected-test scope, or fixed checks only from the base diff. Existing framework changes, governance/shared machinery, removals, untested/multi-module additions, and uncertain executable changes still force full. Integrated full validation is weekly/manual instead of repeated on every PR and merge.

The updated local boundary passed all fixed checks plus 32 focused selector/workflow/policy tests. The selector chooses exactly those four affected test files for #83.

[vantasnerdan]

Live CI proof: Actions run 32122559445 completed in 54 seconds using Validation mode: scoped; it selected the four declared process/selector test files, passed 32 tests plus fixed checks, and skipped the periodic/manual full step. The hard-coded per-PR full replay is removed in #83.

[vantasnerdan]

Process handoff from PR #87

Merged

• Remote validation workflow removal at .github/workflows/validate.yml — removes the third replay of the same repository scripts.
• Local validation contract alignment in AGENTS.md, AGENTS_START_HERE.md, the PR template, research-pr-harvest skill, and memory templates — submitting and reviewing agents retain proportionate local evidence without a mandatory GitHub Actions replay.
• Local selector and regression alignment in scripts/validate_changed.py and tests/test_public_contribution_surfaces.py — the selector remains available and the deleted workflow is no longer required.
• Main branch protection update — repository-validation is no longer a required status; review, code-owner, last-push approval, conversation resolution, admin enforcement, force-push prohibition, and deletion prohibition remain.

Evidence: 20 directly affected regressions passed; GitNexus found zero indexed symbols or execution flows; diff check clean. Landed merge: 1292588.

Requires refactor

• None.

Left in PR history

• The deletion-only first commit — locally correct but incomplete until normative references and fixed tests were repaired by 6a23c65.

Continuation

• Claims promoted: none.
• Goal state: open.
• PR lifecycle: merged; exact head branch deleted.
• Next decisive action: apply the corrected advancement-oriented lifecycle to the ready scientific promotion PRs without recreating duplicate remote validation.
• Terminal-close evidence: not applicable while issue [Process] Keep viable scientific PRs open and actively harvest or challenge canon #81 remains open.

What closes the goal

• Objective-level closure: the active PR lifecycle consistently advances viable scientific work, preserves finite refactor paths, and reserves closure for explicit terminal dispositions.
• Current-route gap: operational use of the landed process on remaining live PRs.
• Necessary obligation: review and disposition those PRs under the landed lifecycle with unit-level rationale.
• One sufficient next move: complete the distinct review/merge disposition of the ready promotion successor.
• Grounds for reconsideration: evidence that removal of remote replay leaves a failure mode not covered by submitter/reviewer local validation and periodic integrated-main checks.
• Reviewer-role boundary: this PR changed process and repository settings only; it promoted no scientific claim.