idkmesh

PR Backlog Convergence and Real-Node Evidence

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

Project-owner request

Continue. We also have many pull requests.

Interpretation

The repository had reached a form of integration backpressure: candidate work, research branches, planning branches, controller experiments, and evidence paths were being generated faster than the project could integrate or reject them.

The appropriate response was therefore convergence before expansion:

inventory the live PR queue;
merge bounded, current, green, canonical changes;
close stale/superseded/deferred branches with explicit successor paths;
preserve useful ideas as public provenance rather than keeping every branch live;
prioritize the real worker -> independent verifier evidence path;
avoid creating another feature PR during cleanup;
keep the final independent human-review boundary intact.

The first quick count undercounted the backlog; a full refresh showed 24 open PRs. The correction was made immediately and the full queue became the convergence target.

Convergence actions

Merged during this pass

#88 — non-selecting Evidence Report/replay: integrated the replayable evidence-report layer without selection or merge authority.
#106 — clean ACE cohort observer: replaced stale #40 with the current trusted cohort/exposure observer.
#104 — ACE live open-work carrying capacity: replaced irreversible cumulative historical review-load pressure with recoverable live open-work pressure while preserving fail-closed actuation gates.
#99 — current execution target graph + repository improvement loop: landed one current planning surface and absorbed the durable convergence-before-expansion contract from #53.
#112 — fail-closed ACE Phase-B activation gate: landed an offline deterministic guard that remains BLOCKED unless protection, real independently verified descendant evidence, capacity, write-budget, and safety conditions all pass. It adds no GitHub mutation or autonomous merge authority.
#113 — real node -> independent verifier E2E: landed the exact-SHA bridge from the accepted canonical node candidate to the current hardened EvaluatorPlan v0.2 / deterministic unified-diff verifier.

Several other canonical pieces also landed concurrently from other repository work while the convergence pass was running, including the clean evaluator/output-authority and security-hardening path (#103, #107, #111) and the ACE shadow/controller foundation (#68). The queue was refreshed repeatedly rather than acting on stale snapshots.

Closed as stale, superseded, or deliberately deferred

#59 old scheduling implementation — superseded by the landed R2 scheduling/churn/scale program.
#56 large ACO/HSR simulator stack — canonical stigmergy path is now the smaller R4 randomness_lab implementation; unique homeostatic ideas remain candidates for focused policy variants.
#53 separate repository-improvement planning branch — its durable operating contract was absorbed into #99.
#40 old ACE cohort observer — superseded by merged #106.
#55 mixed memory-audit + criticality draft — criticality remains a focused research question rather than a mixed-scope live branch.
#36 stale Repository Homeostasis draft — bounded repository-observatory/migration work remains under the current structural issues instead of a second control plane.
#63 interoperability/A2A/MCP draft — deferred until the local real worker/evidence loop is stable.
#89 old Phase-B activation branch — replaced by the refreshed, fail-closed current-state guard #112.
#44 early ACE toy population simulator — preserved as historical mechanism evidence; current ACE capacity/controller paths supersede it.
#43 old GitHub Reflex Observatory draft — preserved as public observability/self-evolution provenance but removed from active review capacity while the real product evidence path is prioritized.
#108 evaluator-owned runtime evidence branch — used to discover/falsify runtime assumptions, then closed rather than becoming a permanent parallel subsystem after its useful evidence was incorporated into the canonical path.

Real worker evidence: failure first, then pass

The most important outcome of convergence was not simply fewer pull requests. It was that the real runtime gate found a genuine defect that static/schema CI had missed.

First frozen runtime attempt

An older PR #91 candidate:

d638a2f78e4a89353b98e91052233e365f56f90a

reached real Docker execution and correctly resolved the immutable python:3.12-alpine image identity, but the positive smoke task failed with a Python SyntaxError.

Root cause: JSON decoding turned newline escapes inside the python -c smoke command into literal newlines inside a single-quoted Python string.

The gate was not weakened and the failed result was preserved as useful evidence.

Worker correction

The canonical smoke task was changed to an escape-free construction using chr(10). A regression test was then added that parses the actual canonical WorkUnit and compiles the decoded python -c command before runtime acceptance.

This produced the current exact worker candidate:

520ad2c9aa5825476de4957da4702d6823f4edb3

Exact-head prerequisites passed:

Node CI 33185901079 — success;
Phase 0 33185901058 — success.

Controlled Docker acceptance

The full exact-head controlled-runtime matrix then passed for 520ad2c9....

Positive evidence included:

schema-valid succeeded ResultManifest;
only README.md changed;
zero untracked/path/unpackaged-artifact/protected-metadata/output/runtime/aggregate policy violations;
non-truncated candidate patch;
independently matching candidate/stdout/stderr digests;
immutable image ID: sha256:1148adc3b50891ee01e6add44be22d277dedeb8dcac2f7cc8b700dd710c75c4d;
matching repository digest: python@sha256:d09d15e60962ca365d1cd544a48773bac9d33f2fb1b00f2aa0deec78ade7dc31;
network none, read-only root, dropped capabilities, no-new-privileges, PID/CPU/RAM bounds, read-only Git metadata, no Docker socket, immutable image execution;
worker acceptance authority remained false.

Required negative cases all failed closed:

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

Issue #37 was completed for the exact corrected worker head.

Evaluator evidence: another useful failure

The first attempt to connect the real accepted worker to the canonical independent verifier exposed a second integration defect, this time in the evaluator-owned harness rather than the worker.

The stale E2E plan failed schema validation because current EvaluatorPlan v0.2 requires:

backend.required_log_types;
deterministic patch-verifier adapter version 0.1.1.

The worker was not modified. The canonical evaluator had already been hardened in #111 to require explicit stdout/stderr evidence and conservative unified-diff structure.

A clean replacement E2E in #113 fixed plan drift by deriving from the current canonical evaluator plan fixture instead of hardcoding evolving evaluator-version details.

Real node -> independent verifier E2E

PR #113 exact-head workflow run 33186566651 passed.

Observed evidence:

accepted worker candidate: 520ad2c9aa5825476de4957da4702d6823f4edb3;
candidate code executed by verifier: false;
canonical patch negative matrix: passed;
WorkUnit: node/canonical-smoke, version 2;
WorkUnit digest: sha256:40993e892a5b83962364686809f7ec6e94ef379e10aaea9492a0526ed7695e2e;
ResultManifest digest: sha256:b45426954a9355629d4746d24ba3b4680962ec96bf7575fd056478a033cfa502;
candidate patch digest: sha256:8383a0dd5217e9472e5f55eb658248620e539394cb96012dc61c24a3cc33f6cf;
EvaluatorPlan digest: sha256:893e59d8d1f8be5bb30e664561eca7bc31d9eb8d3c743225f7e63662b0912c1b;
required evaluator log types: stdout, stderr;
verifier adapter version: 0.1.1;
verifier independent from worker: true;
result-manifest-schema: passed;
independent-review: passed;
VerificationResult status: passed;
VerificationResult digest: sha256:f52686e8e715ecc19ca9788c221d268b4772846aa4a756c18a43ebbf952711cd;
recommendation: accept_candidate as decision support only;
human_integration_decision_required: true.

Issue #5 Phase B1 is therefore complete. The next verification milestone is a small 5–10 task real replayable cohort before any larger benchmark expansion.

Final PR state

At the end of this convergence turn, the active pull-request queue is:

#91 only — canonical local idkmesh-node worker candidate.

This is intentional. #91 remains draft because the project explicitly requires a separate human/reviewer inspection of the exact-head CI and controlled-runtime evidence before integration. It has been labeled help wanted to make that community review need visible.

The low PR count is not permission to weaken the final independence boundary.

Durable operating lesson

IDKMesh should treat PR backlog like verification debt:

proposal generation > review/integration capacity
  -> open-work pressure rises
  -> context and merge drift grow
  -> duplicate mechanisms accumulate
  -> integration quality falls

The corrective controller is:

high open-work pressure
 -> convergence mode
 -> stop optional new branches
 -> merge current bounded evidence
 -> close/defer superseded work
 -> preserve provenance
 -> repair real failures
 -> re-measure the queue

A failed test, runtime gate, or verifier run is not wasted work. If it reveals a real incompatibility and produces a regression guard, it is high-value project evidence.

Community impact

Reducing the PR queue from 24 to 1 makes the repository easier for a newcomer or independent reviewer to understand. The remaining open contribution surface is also unusually concrete: review exact worker head 520ad2c9... and its already-recorded CI/runtime/E2E evidence without needing to reason about two dozen competing branches.

The next product work after that review should remain evidence-first: integrate the canonical worker, run real multi-attempt orchestration through the landed verifier and non-selecting Evidence Report, then build the first small real benchmark cohort.

This site is open source. Improve this page.