idkmesh

Branch and PR convergence continuation

Date: 2026-08-28
Repository: MSKazemi/idkmesh

Owner instruction

Continue maintaining IDKMesh professionally and consider branches as well as pull requests, not only the visible open PR queue.

Maintainer interpretation

A Git branch is not an integration unit by itself.

The repository had accumulated a large branch graph while PRs were being merged, squashed, superseded, rebuilt, used for evidence, and sometimes reused after merge. Treating every ahead/diverged branch as mergeable work would reintroduce duplicate implementations and stale ancestry.

The operating rule for this pass was therefore:

branch -> classify provenance/current value
       -> exact diff/evidence review
       -> clean current-main PR when useful
       -> preserve negative/evidence history
       -> retire only when safe

Direct branch merge authority remained zero throughout.

Canonical branch audit

The repository already contained the read-only branch-convergence auditor from merged PR #123. It was reused rather than replaced.

The auditor:

• reads branches and PR lineage;
• distinguishes squash-integrated source heads from true unique work;
• preserves active/evidence-sensitive branches;
• rejects direct branch merge as an integration strategy;
• has read-only contents / pull-requests permissions.

First refreshed snapshot in this pass

The first live refresh observed:

total branches                         113
cleanup eligible                        59
integrated via merged PR                54
orphan / no unique commits               5
post-merge branch moved                  5
orphan diverged                          7
closed-unmerged evidence branch          6
closed-unmerged unique work             34

That snapshot was used to inspect moved and orphan branches rather than immediately deleting refs.

Settled snapshot after convergence work

Final settled audit:

• workflow run 33187644986;
• job 98927547838;
• artifact 9695119980;
• artifact digest sha256:e9f8358bbd2eece0ed78c0875b469f12ec080cfce52638239073b81b25c9065d.

Result:

branches observed                      131
non-default branches                   130
cleanup eligible                        69
direct branch merges allowed             0

canonical                                1
active draft PR                          1
integrated via merged PR                62
orphan / no unique commits               7
orphan diverged                          7
post-merge branch moved                  5
closed-unmerged evidence branch          9
closed-unmerged unique work              39

The branch count increased because the repository continued creating and merging research/evidence work while maintenance was underway. The important result is not a cosmetically low count; it is an explicit lifecycle state and a safe convergence decision for the high-risk categories.

The canonical maintenance ledger is issue #127.

Post-merge moved branches

A merged PR does not authorize commits added to its branch later. Five such branches were inspected.

acceptance/real-node-verifier-520ad2c

The branch had advanced after merged PR #113 by a missing conversation record. That record was preserved on main as:

docs/conversations/2026-08-28-real-node-verifier-e2e.md

After preservation, the stale source branch has no remaining integration reason.

ace-hardening-converged-v0

The branch was reused after merged #98 for later R3 research already represented by the canonical research lineage. Do not merge the reused branch.

feat/conjunctive-evolution-current-main

Its executable post-merge minimization delta was later represented by merged #150. Do not merge the reused source branch.

fix/resource-registry-runtime-boundary

Later branch movement was represented by the canonical runtime/evidence path. Do not merge stale ancestry.

maintenance/branch-convergence-audit-v0

The branch was reused after its audit PR for later Free Resource / benchmark-contract work that landed through separate reviewed PRs. It should not be merged as one combined branch.

Diverged orphan decisions

Retire / do not revive

• feat/local-independent-validator-v2 — duplicate obsolete validator/EvaluatorPlan-v0.1 architecture; canonical verifier/Evaluator Sovereignty supersedes it.
• idkgraph/p1-adr-linkage-triage — useful semantics rebuilt cleanly and merged through #155.
• idkgraph/p1-warning-sample — useful sampler rebuilt and merged through #162, with follow-up classification #166.
• planning/refresh-current-priorities-post-evidence — replacement #154 became known-stale before integration; historical reasoning was preserved but a false CURRENT_PRIORITIES snapshot was not merged.
• integration/real-node-verifier-e2e — old real-node/verifier harness superseded by #108/#113 and later real two-attempt/report/replay evidence.
• planning/current-execution-graph-v2 — operationally stale planning snapshot, superseded by current planning/triage surfaces.
• fix/phase-b2-v03-calibration — historical pre-v0.4 calibration work; semantic boundary superseded by #164/#171 and later Task-001 calibration evidence.

Extract / rebuild, never merge wholesale

interop-runtime-integration contains a useful A2A/MCP binding idea but is more than a hundred commits behind and hard-codes Work Unit schema 0.1 while the project has newer canonical contracts. Issue #17 remains the correct tracker. Only the interop semantics should be rebuilt on current main, with current contract/SDK conformance and heterogeneous adapters.

growth/fast-discovery-20260828 contains useful discovery/first-contact analysis. Its primary current admin action is already durable in issue #173. The stale branch should not be merged wholesale; durable docs/front-door content may be extracted through a clean current-main PR after time-sensitive external free-resource claims are revalidated.

IDKGraph and ACE convergence

PR #155 — merged

Resolved accepted-ADR linkage warnings by adding explicit evidence-backed references, not by weakening detectors.

Exact live observatory result before merge:

• 269 nodes;
• 0 deterministic errors;
• 0 accepted-decision linkage warnings;
• residual orphan-document findings remained warnings for later review.

PR #156 — merged

Rebuilt the ACE population experiment on canonical recoverable live-open-work-v1 capacity instead of the obsolete cumulative-load assumption. The experiment remains deterministic/offline/toy research and creates no Phase-B actuation authority.

PR #162 / #166 — merged concurrently

Rebuilt the orphan-warning sampler on current main and then classified a frozen warning cohort without weakening IDKGraph semantics.

Phase B2 benchmark falsification and semantic versioning

The benchmark lane produced important negative evidence rather than being forced to pass.

Task 001 evidence failure

PR #158 generated a straightforward Task-001 fix that passed its path-boundary behavior but was rejected by the frozen metadata evaluator. The mismatch was real: the frozen plan used a semantic fragment while verifier 0.1.1 interpreted required_added_text as exact whole-line equality.

The correct response was not to edit the frozen benchmark after seeing the answer.

Burn the pilot

Merged PR #160 burned the first-five pilot while preserving the original frozen WorkUnits, EvaluatorPlans, and pre-outcome definition digest.

Explicit semantic boundary

PR #164 introduced v0.3 added-line substring semantics while preserving v0.2 exact-line meaning.

Pre-v0.3 successor-freeze PR #163 was closed instead of silently translating its frozen commitment; its anti-Goodhart reasoning was preserved in:

docs/conversations/2026-08-28-phase-b2-evaluator-freeze-audit-and-v2.md

v0.4 transition semantics

Parallel #170 and #171 initially introduced overlapping verifier-0.3.0 implementations.

#171 was selected as the canonical layer because it cleanly preserves:

v0.2 / 0.1.1 = exact added-line equality
v0.3 / 0.2.0 = added-line substring matching
v0.4 / 0.3.0 = required added + removed line substring transitions

Its exact-head Evaluator Plan Binding, Phase 0, randomness, IDKGraph, and Evolution checks were green. It merged as commit:

c60549c43232231c724fe3aaaac1f08a26998cbe

#170’s earlier Task-001 calibration run failed because its supposed inert decoy unexpectedly changed the vulnerable behavior. The branch subsequently advanced and its latest head produced a successful calibration run. That calibration evidence is useful, but #170 remains closed because it also contains a second v0.4 schema/runner/verifier implementation. The correct next action is to extract only the successful Task-001 behavioral/adversarial calibration onto current main, calling the canonical #171 transition verifier.

Issue #157 records this remaining calibration extraction boundary.

Canonical node branch convergence

Historical PR #91 contained the accepted worker bytes but had become a stale integration branch relative to modern main.

Clean replacement PR #159 was created from current main.

At convergence review:

• historical #91 head: 520ad2c9aa5825476de4957da4702d6823f4edb3;
• replacement #159 head: 61cafa86f7e0e86343d73182862e3cead1080ab9.

Git-tree/blob inspection confirmed all 14 worker/workflow/test/schema blobs carried by #91 were byte-identical on #159. #159 adds only a convergence conversation record.

Byte equality was explicitly not treated as acceptance equality.

#91 was closed as the historical exact-evidence branch so the repository would have one canonical worker integration path.

Historical #91/#37/#108 evidence remains public provenance for the old exact head, including the failed-first Docker run, fixture correction, regression guard, and later passing matrix.

Fresh #159 evidence

Fresh exact-head CI on #159:

• Node CI 33193136252 — success;
• Phase 0 33193136271 — success;
• randomness-lab 33193136289 — success;
• Evolution 33193136417 — success.

A new candidate-bound read-only evidence PR #169 reused the previously reviewed #108 runtime acceptance mechanics by blob identity and changed only the replacement candidate/CI binding.

#169 workflow:

• run 33193838388;
• job 98925820770;
• candidate 61cafa86f7e0e86343d73182862e3cead1080ab9;
• all_acceptance_checks_passed: true;
• worker_acceptance_authority: false.

Positive evidence confirmed:

• schema-valid succeeded ResultManifest;
• expected README.md-only candidate change;
• no path/untracked/unpackaged/protected-metadata/output/runtime violations;
• non-truncated patch;
• matching patch/stdout/stderr digests;
• immutable image ID + matching repository digest;
• network-none, read-only root, capability drop, no-new-privileges, PID/CPU/RAM bounds, read-only Git metadata, and no Docker socket.

Negative A-E2 cases all failed closed:

• forbidden/out-of-scope tracked path;
• ignored untracked artifact;
• task-visible Git pointer tampering;
• oversized/truncated patch;
• absent local image;
• locally retagged image without matching repository digest.

#169 was then closed without merge because it is one-shot candidate-bound evidence instrumentation. Its run/logs remain provenance without leaving a stale SHA-bound workflow on main.

Remaining worker gate

PR #159 is intentionally still draft.

Mechanical runtime acceptance is complete. The remaining gate is a genuinely separate human/reviewer inspection through issue #138.

Neither blob equality, worker success, green CI, the runtime harness, the PR author, nor project automation is allowed to manufacture that approval.

If #159’s head/tree changes, the relevant evidence must be rebound/re-evaluated.

Branch deletion boundary

The connected GitHub maintenance surface available in this session does not expose branch-ref deletion.

Therefore this pass did not fake deletion by force-moving refs.

Issue #127 records 69 cleanup-eligible branches at the settled audit snapshot. A deletion-capable repository-admin pass should:

1. revalidate each exact branch head immediately before deletion;
2. preserve evidence-sensitive refs until durable run/commit/artifact references are sufficient;
3. confirm no open PR or external workflow depends on the branch name;
4. delete only verified cleanup-safe refs;
5. rerun the canonical audit and measure the actual reduction.

Branch count is coordination hygiene, not a project-success metric.

External integration and growth gates

Public metadata continued to report main as unprotected. Issue #35 remains the P0 repository-admin safety gate before stronger autonomous repository writes.

Public repository metadata also still showed an empty description, no topics, Discussions disabled, Pages disabled, homepage unset, and zero stars/forks. Issue #173 tracks the GitHub-native discovery/admin actions. These settings should be improved without turning stars/comments/PR count into the optimization target.

End state of this continuation

At the final settled branch audit:

131 branches
69 cleanup-eligible
0 direct branch merges allowed
1 active draft PR

After resolving the overlapping v0.4 review queue, the live open-PR search again contained only #159.

The next highest-value actions are therefore external or evidence-bound rather than another parallel architecture branch:

1. separate human review of #159 (#138);
2. protect main (#35);
3. physically retire exact cleanup-safe branches through a deletion-capable admin surface (#127);
4. extract the successful Task-001 behavioral calibration onto canonical v0.4 without a second verifier implementation (#157);
5. rebuild interoperability on current contracts rather than reviving interop-runtime-integration (#17);
6. activate GitHub-native discovery surfaces (#173).

The repository should continue preferring convergence, independent evidence, and clear contributor entry points over raw branch/PR/event volume.

This site is open source. Improve this page.