12.20 Issue 61 NNx Release Review Implementation Plan

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.

Goal: Decide with released-wheel, consumer, notebook, security, and GitFlow evidence whether ml-eng-lab can adopt thekaveh-nnx[lm]==0.2.2; retain that pin only when every mandatory contract passes.

Architecture: Keep requirements.txt as the single production version source and keep the Issue #58 verifier as the canonical provenance oracle. Strengthen all NNx-executing CI tiers to install the binary distribution and verify it immediately before their workload, add focused consumer tests for the newly relevant 0.2.2 surface, refresh the advisory observation without weakening its accepted baseline, and project one canonical documentation truth to the repository, site, and wiki.

Tech Stack: Python 3.11, pytest, PyTorch 2.4.1, PyG, thekaveh-nnx[lm], papermill, GitHub Actions, pip-audit, MkDocs Material, GitHub wiki, Git, and GitHub CLI.

Spec: docs/superpowers/specs/2026-08-13-issue-61-nnx-release-review-design.md

12.20.1 Global Constraints

• Work only on codex/issue-61-nnx-release-review, created from the synchronized develop branch. Preserve unrelated user changes if any appear.
• Use test-driven development: add a focused failing test or mutation first, capture the intended RED, make the smallest production change, and rerun the focused GREEN gate before broad tests.
• Keep requirements.txt as the only production NNx version source. Do not add a candidate manifest, version constant, compatibility flag, editable bypass, direct URL, or downstream shim.
• Require python -m pip install --only-binary=thekaveh-nnx -r requirements.txt wherever CI claims NNx released-wheel evidence. Run make verify-nnx-install after the final install and immediately before the NNx workload.
• Do not weaken scripts/verify_nnx_install.py, workflow failure propagation, notebook acceptance thresholds, audit comparison, provenance checks, or skip guards to make 0.2.2 pass.
• Treat the active editable checkout as development context only. It is not release evidence even if its source reports 0.2.2. Use disposable clean environments and GitHub-hosted jobs for acceptance.
• Keep the root Torch/PyG/Lightning/torchao pins unchanged. Issue #62 owns their coordinated upgrade; Issue #66 owns restored quantization execution and tiering.
• Do not edit the infra/ gitlink or Atlas-owned requirements. Preserve the observed Atlas NNx 0.2.0 evidence as historical runtime truth. Do not launch Atlas, JupyterHub, Docker Compose, Ollama, or ComfyUI for this issue.
• Keep the active tabular-regression notebook's manual loader and existing results. Document that 0.2.2 now offers target_dtype=torch.float32 without changing the experiment split.
• Do not add NNGraphDataset(seed=...) to locked notebooks, migrate NNDataset(batch_sizes=...), replace deep imports, change the MoE architecture, or introduce a Conv notebook. Issues #68, #69, and #70 own those changes.
• Preserve synthetic 0.2.0 values where they are generic verifier fixtures and preserve all historical Issue #58, changelog, maintenance, and Atlas evidence. Change only current-pin and current-user-contract references.
• Never auto-accept a new advisory or accepted-version drift. A new identity, changed accepted version, unexplained surface drift, or failed audit blocks adoption and requires explicit security review.
• Edit canonical documentation only. Regenerate and inspect generated/site and generated/wiki, but do not commit either ignored tree.
• Use apply_patch for tracked text edits. Use repository formatters or provided scripts only for mechanical formatting or generated derivatives.
• Every task receives an implementation review. After all tasks, request an independent broad review and fix every Critical, Important, and Minor finding with fresh RED-to-GREEN evidence.
• Merge through a ready feature-to-develop PR and a separate develop-to-main PR. Run the labeled Tier B job and manually dispatched Tier C job on the feature ref before the first merge.
• Finish with main/develop synchronization, Issue #61 and parent #53 bookkeeping, project-board completion, ruleset readback, and local/remote branch, PR, worktree, and repository-container cleanup.

12.20.2 File and Interface Map

12.20.2.1 Production and workflow contracts

• Modify requirements.txt: change the single root pin to thekaveh-nnx[lm]==0.2.2 and retain its binary-wheel ownership commentary.
• Modify Makefile: update only current release commentary; preserve the version-deriving verify-nnx-install target and all tier membership.
• Modify .github/workflows/ci.yml: strengthen tier-a-papermill, smoke-tier-b, and smoke-tier-c with binary-only selection and canonical verification. Preserve triggers, conditions, job names, permissions, timeouts, artifacts, and required contexts.
• Do not modify scripts/verify_nnx_install.py unless a test demonstrates a version-independent defect. Its public contract remains parse_requirement, discover_distribution, and fail-closed main verification derived from requirements.txt.

12.20.2.2 Consumer tests

• Modify tests/test_verify_nnx_install.py: add or update only the current-repository 0.2.2 expectation; retain generic 0.2.0 mismatch fixtures.
• Modify tests/test_verify_repo.py: encode exact Tier A/B/C install, provenance, ordering, failure-propagation, and no-override workflow contracts with isolated mutations.
• Create tests/nnx_surface/test_tabular_regression_diabetes_mlp_pytorch.py: validate floating regression targets, (batch, 1) shape, output_dim == 1, default classification behavior, and a tiny MSE training pass.
• Create tests/nnx_surface/test_nnx_022_release_surface.py: validate the public Conv and first-class MoE facade, constructor acceptance, and stable enum members without importing deep implementation modules.
• Modify tests/nnx_surface/test_notebook_api_surface.py: recognize released NNGraphDataset(seed=..., sampler=...) kwargs while leaving notebook source unchanged.
• Modify tests/nnx_surface/test_quantization_mnist_ffnn_pytorch.py only if the existing exact QAT signatures do not already prove the 0.2.2 facade. Keep the Torch 2.4.1 backend skip honest.

12.20.2.3 Current documentation truth

• Modify README.md, CONTRIBUTING.md, CHANGELOG.md, and docs/assets/badges/nnx.svg for the current 0.2.2 adoption and immutable wheel/license evidence. Do not rewrite historical entries.
• Modify docs/nnx-library.md, docs/dependency-contracts.md, docs/architecture.md, docs/FINDINGS-NNX.md, and docs/concepts.md for the current release, evidence boundary, resolved regression support, and valid checkpoint contract.
• Modify docs/notebooks/tabular_regression-diabetes-mlp-pytorch.md, notebooks/tabular_regression-diabetes-mlp-pytorch/README.md, notebooks/tabular_regression-diabetes-mlp-pytorch/docs/spec.yaml, and only the relevant Markdown cell in notebooks/tabular_regression-diabetes-mlp-pytorch/notebook.ipynb. Preserve code, output, split, and recorded metrics.
• Modify current-pin references in docs/notebooks/text_generation-tinyshakespeare-transformer-pytorch.md, notebooks/text_generation-tinyshakespeare-transformer-pytorch/README.md, and notebooks/preference_alignment-toy-dpo-pytorch/README.md.
• Reconcile every current canonical NNRun.load("best") claim, including the Iris and dimensionality-reduction canonical pages and Iris task spec, to NNCheckpoint.load(run=RUN_ID, type=Checkpoints.BEST). Preserve historical descriptions of the former bug.
• Modify tests/test_check_docs.py to enforce current pin parity, release-boundary wording, regression-finding status, valid checkpoint wording, and consecutive manifest entries 12.19 and 12.20.
• Modify docs/manifest.yaml only in the plan commit to register this page as 12.20.

12.20.2.4 Security evidence and rollout

• Refresh the current observation in docs/dependency-contracts.md only after the exact make audit-advisories result is captured. Do not change security/accepted-advisories.json unless the controller separately approves a reviewed policy decision.
• Update .github/workflows/docs.yml paths only if a newly changed canonical source is not already covered. Do not add redundant literals beneath the existing root *.md coverage.
• Record implementation evidence in ignored .superpowers/sdd/ reports; never stage those reports.

12.20.3 Task 1: Adopt the Canonical 0.2.2 Pin and Strengthen Tier Provenance

Files: requirements.txt, Makefile, .github/workflows/ci.yml, tests/test_verify_nnx_install.py, tests/test_verify_repo.py

• [x] Step 1: Revalidate the release before editing

Query PyPI and the immutable GitHub release for 0.2.2. Confirm it is still the latest stable, non-yanked universal wheel; confirm tag commit edfd197f3f54d4eb67313d46a80e823e6239c5b6, wheel SHA-256 ee56474926fdfd5329721f067cf1b8ae31955627c6949844e09ee4a7bb2bb9d7, [lm] extra, Python/Torch/PyG floors, and Apache-2.0 license. Stop on any mismatch.

• [x] Step 2: Write current-pin and Tier A/B/C workflow RED tests

Require the root requirement to parse as exactly thekaveh-nnx[lm]==0.2.2. For each of tier-a-papermill, smoke-tier-b, and smoke-tier-c, require this ordered boundary:

make install-torch-stack
python -m pip install --only-binary=thekaveh-nnx -r requirements.txt
make verify-nnx-install
make smoke-tier-a  # or the exact smoke-tier-b / smoke-tier-c workload for that job

Tier A may retain its two asset downloads between dependency installation and provenance, but no package installation may occur after provenance. Require no editable/path/direct-URL override, PYTHONPATH, NNX_ALLOW_EDITABLE, continue-on-error, job container, or service. Require the verifier step immediately before the tier workload.

• [x] Step 3: Run focused tests and capture RED

pytest -p no:cacheprovider tests/test_verify_nnx_install.py tests/test_verify_repo.py -q \
  -k 'current_requirement or tier_a_nnx or tier_b_nnx or tier_c_nnx'

Expected: the current pin remains 0.2.0 and all three tier jobs lack the binary/provenance contract.

• [x] Step 4: Apply the minimal pin and workflow change

Change the sole root pin to 0.2.2. Replace each tier job's plain runtime install with the exact binary-only command, add a named canonical-verifier step, and make it the final step before the workload. Do not modify verifier implementation or unrelated jobs.

• [x] Step 5: Mutation-audit the workflow contract

In isolated workflow copies, prove RED for removed verifier, reversed verifier/workload order, plain pip install, editable install, late install, alternate requirements file, job-level PYTHONPATH, NNX_ALLOW_EDITABLE, failure masking, extra Atlas/service steps, and renamed workload. Restore the clean workflow after every mutation.

• [x] Step 6: Run Task 1 GREEN gates

pytest -p no:cacheprovider tests/test_verify_nnx_install.py tests/test_verify_repo.py -q \
  -k 'current_requirement or tier_a_nnx or tier_b_nnx or tier_c_nnx'
python -c "from pathlib import Path; import yaml; yaml.safe_load(Path('.github/workflows/ci.yml').read_text())"
ruff check tests/test_verify_nnx_install.py tests/test_verify_repo.py
git diff --check

• [x] Step 7: Commit Task 1

git add requirements.txt Makefile .github/workflows/ci.yml \
  tests/test_verify_nnx_install.py tests/test_verify_repo.py
git commit -m "deps: trial NNx 0.2.2 canonical wheel"

12.20.4 Task 2: Add Consumer-Facing 0.2.2 Compatibility Coverage

Files: tests/nnx_surface/test_tabular_regression_diabetes_mlp_pytorch.py, tests/nnx_surface/test_nnx_022_release_surface.py, tests/nnx_surface/test_notebook_api_surface.py, and, only if needed, tests/nnx_surface/test_quantization_mnist_ffnn_pytorch.py

• [x] Step 1: Write the regression-dataset contract

Build a deterministic tiny pandas frame and require:

regression = NNTabularDataset(..., target_dtype=torch.float32)
assert regression.output_dim == 1
assert target.dtype == torch.float32
assert target.shape[1:] == (1,)

Run one tiny MSE optimization through the public NNx facade and assert finite loss/prediction shape. Add a paired default-None test proving Iris-style integer class targets remain contiguous torch.long values and retain multi-class output_dim.

• [x] Step 2: Write additive facade contracts

Import ConvNN, NNConvParams, FeedFwdMoENN, NNMoEParams, and Nets from nnx only. Construct the smallest valid Conv and MoE configurations, assert Nets.CONV and Nets.FEED_FWD_MOE, and execute one shape-level forward pass where the public model contract permits it. Do not migrate any notebook to these architectures.

• [x] Step 3: Extend the graph signature guard

Add synthetic accepted calls with seed=0 and sampler="full" and make the accepted-signature fixture include both names. Do not edit Reddit notebooks; the live signature-derived scan must continue to validate their existing kwargs.

• [x] Step 4: Capture release-delta RED in a disposable 0.2.0 environment

Create a temporary Python 3.11 virtual environment outside the repository, install the released 0.2.0 binary wheel plus only the focused test dependencies, and run the new release-surface tests. Expected RED: regression target dtype and/or new facade/graph contracts are absent. Delete the disposable environment after recording the sanitized result.

• [x] Step 5: Install and verify the clean 0.2.2 wheel in a disposable environment

Install the repository manifests using the existing Torch-first order and exact binary-only NNx command. Run make verify-nnx-install with the disposable interpreter and record version, non-editable provenance, WHEEL/RECORD ownership, and import origin without disclosing host paths.

• [x] Step 6: Run focused GREEN tests

make verify-nnx-install
pytest -p no:cacheprovider \
  tests/nnx_surface/test_tabular_regression_diabetes_mlp_pytorch.py \
  tests/nnx_surface/test_nnx_022_release_surface.py \
  tests/nnx_surface/test_notebook_api_surface.py \
  tests/nnx_surface/test_quantization_mnist_ffnn_pytorch.py -q
make test-nnx-surface
ruff check tests/nnx_surface
git diff --check

If the host interpreter is editable, run these gates inside the verified disposable wheel environment; do not set NNX_ALLOW_EDITABLE=1 for release evidence.

• [x] Step 7: Commit Task 2

git add tests/nnx_surface/test_tabular_regression_diabetes_mlp_pytorch.py \
  tests/nnx_surface/test_nnx_022_release_surface.py \
  tests/nnx_surface/test_notebook_api_surface.py \
  tests/nnx_surface/test_quantization_mnist_ffnn_pytorch.py
git commit -m "test: cover NNx 0.2.2 consumer surface"

Stage the quantization file only if it changed.

12.20.5 Task 3: Refresh the Fail-Closed Advisory Observation

Files: docs/dependency-contracts.md, tests/test_verify_repo.py, and only if an existing generic parser defect is proven, scripts/verify_repo.py or scripts/advisory_baseline.py

• [x] Step 1: Capture the pre-refresh offline baseline state

Run D10 and record the expected input-hash drift caused by the new requirements.txt. Confirm the accepted-policy file remains canonical and unchanged.

• [x] Step 2: Run the live four-surface audit

make audit-advisories

Capture tool version, UTC timestamp, branch commit, Python/platform, input SHA-256 values, resolved versions, raw feed records, unique identities, and per-surface attribution. Sanitize temporary paths and credentials.

• [x] Step 3: Enforce the adoption stop rule

Compare the new observation with security/accepted-advisories.json and the current ledger. Stop without changing policy if there is any new advisory identity, accepted-version drift, unexplained surface drift, missing required package, invalid schema, resolver failure, or service failure. A disappeared advisory is reconciliation evidence, not proof of remediation.

• [x] Step 4: Write ledger-refresh RED tests

Require the current snapshot to use the fresh timestamp, commit, manifest hashes, result count, and surface rows. Require the historical Issue #59 snapshot and Atlas 0.2.0 observation to remain byte-semantically historical. Add mutations for stale runtime hash/count and accidental Atlas rewrite.

• [x] Step 5: Update only the current human observation

Reconcile docs/dependency-contracts.md with the exact audit JSON. Do not edit security/accepted-advisories.json when identity/version/surface parity is unchanged. Do not claim a package is remediated solely because it is absent.

• [x] Step 6: Run Task 3 GREEN gates

pytest -p no:cacheprovider tests/test_verify_repo.py -q -k 'dependency_ledger or advisory'
python scripts/verify_repo.py --check docs --fast
make audit-advisories
ruff check scripts/verify_repo.py scripts/advisory_baseline.py tests/test_verify_repo.py
git diff --check

• [x] Step 7: Commit Task 3

git add docs/dependency-contracts.md tests/test_verify_repo.py
git commit -m "security: refresh NNx dependency evidence"

Add a production parser file only when its new test proved a version-independent defect. Never stage the accepted-policy JSON without a separate reviewed security decision.

12.20.6 Task 4: Reconcile Current Documentation Across All Three Surfaces

Files: the current documentation files listed in §12.20.2, tests/test_check_docs.py, and docs/manifest.yaml only if a manifest correction is genuinely needed

• [x] Step 1: Write documentation-contract RED tests

Require:

• exactly one current root requirement, thekaveh-nnx[lm]==0.2.2, and agreement among all current user-facing pin references;
• wheel SHA, immutable release/tag evidence, and Apache-2.0 transition on the dependency page;
• current Tier A/B/C canonical-wheel evidence and best-effort quantization boundary;
• the tabular-regression finding marked resolved in 0.2.2 while the manual loader remains an intentional split-preservation choice;
• current user guidance to use NNCheckpoint.load(run=RUN_ID, type=Checkpoints.BEST), with no active NNRun.load("best") claim;
• Atlas image evidence remaining 0.2.0 and explicitly Atlas-owned;
• Issue #58/historical release records remaining historical; and

• consecutive projected design/plan entries 12.19 and 12.20.

• [x] Step 2: Run focused docs tests and capture RED

pytest -p no:cacheprovider tests/test_check_docs.py -q \
  -k 'nnx or manifest or tabular_regression or checkpoint'

• [x] Step 3: Update current release entry points

Update current 0.2.0 references in README, contributor guidance, Makefile commentary, badge, NNx overview, architecture, dependency contracts, TinyShakespeare docs/README, and DPO README. Add one Unreleased changelog entry describing adoption and evidence; do not rewrite older entries.

• [x] Step 4: Resolve tabular-regression prose without changing execution

In the finding ledger, task README, task spec, canonical notebook page, and the notebook's relevant Markdown cell, state that 0.2.2 supports floating regression targets through target_dtype=torch.float32. State that this notebook intentionally retains manual loaders to reuse the established sklearn/NNx split. Do not change code, execution counts, outputs, or metrics.

• [x] Step 5: Correct every current checkpoint claim

Replace active NNRun.load("best") guidance in NNx overview, concepts, Iris canonical page, dimensionality-reduction canonical page, and Iris spec with the valid checkpoint API. Preserve historical changelog and maintenance references describing the prior defect.

• [x] Step 6: Regenerate and inspect all three surfaces

make docs-check
make docs-wiki

Inspect the generated site and wiki copies of the NNx overview, dependency ledger, finding ledger, and tabular-regression page. Confirm current markers and links are equivalent after transformation, design/plan nav entries are consecutive, and generated trees remain ignored.

• [x] Step 7: Run Task 4 GREEN gates

pytest -p no:cacheprovider tests/test_check_docs.py tests/test_build_docs.py \
  tests/test_wiki.py tests/test_manifest.py tests/test_transforms.py -q
make docs-check
make docs-wiki
python scripts/verify_repo.py --check docs --fast
ruff check .
git diff --check

• [x] Step 8: Commit Task 4

Stage the exact changed canonical files and tests, then:

git commit -m "docs: document NNx 0.2.2 adoption"

12.20.7 Task 5: Run the Complete Release-Acceptance Matrix

Files: no intended tracked changes; write sanitized ignored evidence under .superpowers/sdd/

• [x] Step 1: Establish a clean released-wheel environment

From a disposable Python 3.11 environment, install the pinned Torch stack, runtime manifest, and docs manifest with the repository's documented order. Require exact NNx 0.2.2, non-editable/non-direct-url metadata, and canonical wheel ownership before any acceptance test.

• [x] Step 2: Run focused and complete test gates

make verify-nnx-install
make test-nnx-surface
make test
make verify
make lint
make docs-check
make docs-wiki
git diff --check

Record exact pass/skip/warning counts. Only documented Torch 2.4.1 torchao backend skips and platform-specific PyG loader skips may remain, and every skip must execute its guard.

• [x] Step 3: Run Tier A in full

make smoke-tier-a
make check-tier-a-artifacts
make check-tier-a-clean

Require all 18 Tier A outputs, including 17 NNx consumers plus the NumPy control. Confirm source notebooks and tracked files remain unchanged.

• [x] Step 4: Run Tier B and Tier C in full

make smoke-tier-b
make smoke-tier-c

Require all six Tier B and all four Tier C outputs in temporary directories. Do not overwrite committed notebooks or adopt graph seed changes.

• [x] Step 5: Run the best-effort QAT checkpoint probe

In a separate disposable Python 3.11 environment with Torch 2.11.0, torchvision 0.26.0, and torchao 0.18.0, install the 0.2.2 binary wheel and run a tiny QAT prepare/train/convert/LAST-checkpoint reload through NNModel.from_checkpoint, then verify prediction shape/type. Record inability to create the environment as a concern; do not alter root manifests or weaken the main acceptance matrix.

The 0.2.2 trial reconstructed the converted model with exact prediction parity. The retained 0.2.0 fallback also completed prepare/train/convert/LAST reload and inference, but its pre-on_train_end LAST checkpoint has no transform metadata and reloads ordinary linear modules; that non-blocking fidelity limitation is recorded in the ignored acceptance report.

• [x] Step 6: Request task-level and broad reviews

Have fresh reviewers compare every task commit with this plan and the approved design. Then request an independent broad review of the entire branch, including mutation resistance, released-wheel provenance, notebook semantics, security parity, and three-surface docs.

The broad review returned Changes Required: the 0.2.2-only notebook calls conflicted with the declared Atlas-default runtime, and it identified documentation, audit-currentness, QAT-version, CI service-isolation, and rollback-boundary corrections.

• [x] Step 7: Fix every review finding with TDD

Validate each finding, add a failing regression or mutation, apply the narrow fix, rerun focused and affected broad gates, and commit separately. Repeat review until no Critical, Important, or Minor finding remains.

This correction wave restores the 0.2.0 default-runtime contract, removes all five trial-only salts and other 0.2.2-only consumer tests, restores the pre-trial current advisory snapshot, strengthens no-service workflow mutations, and reconciles canonical documentation. Focused, affected, strict documentation, verifier, lint, clean-wheel 0.2.0 surface, and advisory comparator gates passed. The controller owns the final complete 0.2.0 matrix and independent re-review for the Step 8 decision.

• [x] Step 8: Decide adopt or retain

Retain 0.2.2 only when all mandatory local gates and task reviews pass. If a blocker is inherent to 0.2.2, revert the root pin and only release-coupled current changes, reinstall 0.2.0, rerun the complete matrix, and publish the precise incompatibility. Never partially retain new-only consumer calls under a reverted pin.

Final decision: retain 0.2.0. The clean local 0.2.2 trial passed every repository and notebook gate but required five NNModel.train(..., salt=...) calls that the unchanged recommended Atlas JupyterHub runtime cannot execute. After removing those trial-only calls and restoring the Atlas-compatible root pin, a fresh canonical 0.2.0 environment passed Step 2, Tier A 18/18, Tier B 6/6, and Tier C 4/4 with clean source and artifact checks.

12.20.8 Task 6: GitFlow Rollout, Bookkeeping, and Cleanup

Files: GitHub pull requests, Issue #61, parent Issue #53, project board, and local/remote Git state; no intended product-file changes

• [ ] Step 1: Push the verified feature branch

Confirm a clean tracked tree, exact commit scope, no generated docs, no ignored evidence staged, no unintended gitlink change, and no repository service containers. Push codex/issue-61-nnx-release-review to origin.

• [ ] Step 2: Open the ready feature-to-develop PR

Create a non-draft PR with Issue #61 linkage, adopt-or-retain decision, wheel provenance, security result, local test/tier matrix, rollback boundary, and explicit statement that Atlas was not launched. Add the tier-b-smoke label before waiting for checks.

• [ ] Step 3: Dispatch Tier C on the feature ref

Run .github/workflows/ci.yml through workflow_dispatch for the exact feature-branch SHA. Wait for smoke-tier-c and record its run/job URL and conclusion. Reject evidence from a stale SHA.

• [ ] Step 4: Wait for and review all feature PR checks

Require the protected contexts pytest-repository, atlas-consumer-policy, and dependency-audit; focused pytest-nnx-surface; Tier A; labeled Tier B; manually dispatched Tier C; docs; and all other triggered jobs to conclude successfully. Inspect failures rather than rerunning blindly. Address every review thread.

• [ ] Step 5: Read back the active ruleset and merge into develop

Verify the gitflow ruleset still protects main and develop and requires exactly the three intended contexts. Merge the ready PR through GitHub, record its URL and merge SHA, fetch, and fast-forward local develop.

• [ ] Step 6: Open and merge the develop-to-main release PR

Confirm develop contains only intended post-main commits, create a separate ready PR into main, wait for the full required and triggered matrix, resolve every thread, merge, and record the release PR URL and main merge SHA.

• [ ] Step 7: Synchronize main back into develop

If the main merge commit is not already an ancestor of develop, create the normal main-to-develop sync PR, wait for required checks, and merge it. Prove git diff origin/main..origin/develop is empty and main is an ancestor of develop.

• [ ] Step 8: Complete issue and project bookkeeping

Check every Issue #61 acceptance item with exact PR/run/commit evidence, post the closeout comment, close Issue #61, move its project item to Done, and check Issue #61 in parent #53. Do not mark #53 complete while later backlog issues remain.

• [ ] Step 9: Clean local and remote state

Delete the merged feature branch locally and remotely, remove merged temporary sync branches, prune remote refs, close or delete dangling PRs, remove stale worktrees, and prove only intended long-lived refs remain. Do not delete unrelated user branches or containers.

• [ ] Step 10: Prove the final state

Record:

git status --short --branch
git branch -a
git worktree list
git diff origin/main..origin/develop
git merge-base --is-ancestor origin/main origin/develop
gh pr list --state open
docker ps --format '{{.Names}} {{.Status}}'

Confirm no ml-eng-lab Atlas, JupyterHub, Ollama, ComfyUI, or issue-specific container remains. Leave unrelated project containers untouched.

12.20.9 Completion Criteria

• [ ] The latest stable release was revalidated and the adopted or retained version is supported by immutable upstream evidence.
• [ ] One exact root NNx pin owns the production version and every NNx-executing CI tier proves the released binary wheel immediately before its workload.
• [ ] The final consumer suite proves the Atlas-default 0.2.0 notebooks use no new-only training keywords; the removed regression, Conv/MoE, graph-signature, identity-salt, and QAT probes remain recorded as 0.2.2 trial evidence rather than current contracts.
• [ ] Complete tests plus Tier A, Tier B, and Tier C pass in a canonical released-wheel environment; QAT side-environment evidence or its explicit limitation is recorded.
• [ ] The live advisory observation reconciles with the accepted policy without automatic risk acceptance.
• [ ] Current repository, site, and wiki documentation agree on the reviewed 0.2.2 trial, retained 0.2.0 pin, still-open current regression limitation, checkpoint API, evidence boundary, and historical Atlas ownership.
• [ ] Independent task and broad reviews report no unresolved Critical, Important, or Minor findings.
• [ ] Feature-to-develop, develop-to-main, and any required main-to-develop synchronization are merged with green checks and the active ruleset remains correct.
• [ ] Issue #61, project board, and parent #53 bookkeeping is complete.
• [ ] Local and remote branches, PRs, worktrees, generated outputs, and repository service containers are clean.