Skip to content

12.18 Issue 60 Machine-Readable Advisory Baseline 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: Enforce the reviewed dependency-advisory policy with a deterministic JSON baseline, a fail-closed four-surface live comparator, an offline JSON-to-ledger parity check, and a dedicated required GitHub Actions context.

Architecture: scripts/advisory_baseline.py owns canonical JSON parsing, pip-audit normalization, four-surface execution, and live comparison. scripts/verify_repo.py remains offline and enforces parity between security/accepted-advisories.json and the current Markdown ledger. An unconditional dependency-audit job runs exact make audit-advisories and becomes the third required ruleset context after its first live PR run.

Tech Stack: Python 3.11 standard library, pip-audit==2.10.0, JSON, pytest, GNU Make, GitHub Actions, existing MkDocs/wiki documentation pipeline.

12.18.1 Global Constraints

  • The accepted identity key is exactly (normalized package, primary advisory ID, accepted version).
  • The baseline contains 21 unique identities; the Markdown ledger continues to preserve 23 raw feed records.
  • The canonical surfaces, in order, are combined-runtime, torch, documentation, and atlas-contract.
  • New identities, accepted-version drift, package/surface drift, invalid observations, and policy/ledger drift fail closed.
  • A disappeared accepted identity produces a reconciliation notice and does not fail the live comparison.
  • Aliases, fix versions, raw-record duplication, rationale, and history remain human-ledger evidence rather than baseline keys.
  • Never use --ignore-vuln, broad suppression, automatic acceptance, or count-only comparison.
  • make verify must remain offline and network-independent.
  • Pin only the direct audit tool at pip-audit==2.10.0; Issue #63 owns complete dependency locking.
  • Do not change NNx, Torch, PyG, Lightning, torchao, documentation, or Atlas dependency versions.
  • The Atlas audit covers only atlas-contract-requirements.txt; do not initialize or start Atlas or any service.
  • Never start containerized Ollama. No Atlas, JupyterHub, Docker, Docker Compose, Ollama, or ComfyUI operation is required.
  • Preserve the existing pytest-repository and atlas-consumer-policy required contexts and every other ruleset parameter.
  • Use TDD: capture the intended RED before each production change, then the focused GREEN.
  • Do not commit generated site/, generated/site/, or generated/wiki/ output.

12.18.2 Task 1: Canonical policy and live comparator

Files:

  • Create: security/accepted-advisories.json
  • Create: vulnerability-audit-requirements.txt
  • Create: scripts/advisory_baseline.py
  • Create: tests/test_advisory_baseline.py
  • Modify: Makefile
  • Modify: tests/test_makefile_contract.py

Interfaces:

  • Produces Baseline, AcceptedAdvisory, Observation, Comparison, and AuditSurface immutable dataclasses.
  • Produces load_baseline(path: Path) -> Baseline.
  • Produces normalize_pip_audit(surface: str, payload: object) -> Observation.
  • Produces compare_baseline(baseline: Baseline, observations: Sequence[Observation]) -> Comparison.
  • Produces run_audit_surfaces(repo: Path, runner: AuditRunner = subprocess.run) -> tuple[Observation, ...].
  • Produces CLI python -m scripts.advisory_baseline --repo-root PATH and Make target audit-advisories.
  • Task 2 consumes load_baseline, canonical identity/surface helpers, and the baseline dataclasses.

  • Task 3 consumes the exact Make target and tool manifest.

  • [x] Step 1: Write canonical-schema RED tests

Create tests that write temporary JSON and assert:

def test_load_baseline_accepts_canonical_policy(tmp_path): ...
def test_load_baseline_rejects_unknown_root_and_item_keys(tmp_path): ...
def test_load_baseline_rejects_unsupported_schema_version(tmp_path): ...
def test_load_baseline_rejects_missing_extra_duplicate_or_reordered_surfaces(tmp_path): ...
def test_load_baseline_rejects_duplicate_identity_after_pep503_normalization(tmp_path): ...
def test_load_baseline_rejects_unknown_or_duplicate_item_surfaces(tmp_path): ...
def test_load_baseline_rejects_unsorted_items_and_noncanonical_serialization(tmp_path): ...
def test_real_baseline_contains_21_unique_reviewed_identities(): ...

The real-identity assertion must compare exactly this set:

{
    ("pytorch-lightning", "PYSEC-2026-3043", "2.4.0"),
    ("torch", "CVE-2025-2148", "2.4.1"),
    ("torch", "CVE-2025-2149", "2.4.1"),
    ("torch", "CVE-2025-2998", "2.4.1"),
    ("torch", "CVE-2025-2999", "2.4.1"),
    ("torch", "CVE-2025-3001", "2.4.1"),
    ("torch", "PYSEC-2024-259", "2.4.1"),
    ("torch", "PYSEC-2025-191", "2.4.1"),
    ("torch", "PYSEC-2025-194", "2.4.1"),
    ("torch", "PYSEC-2025-198", "2.4.1"),
    ("torch", "PYSEC-2025-203", "2.4.1"),
    ("torch", "PYSEC-2025-204", "2.4.1"),
    ("torch", "PYSEC-2025-205", "2.4.1"),
    ("torch", "PYSEC-2025-206", "2.4.1"),
    ("torch", "PYSEC-2025-207", "2.4.1"),
    ("torch", "PYSEC-2025-208", "2.4.1"),
    ("torch", "PYSEC-2025-209", "2.4.1"),
    ("torch", "PYSEC-2025-41", "2.4.1"),
    ("torch", "PYSEC-2026-139", "2.4.1"),
    ("torch", "PYSEC-2026-1970", "2.4.1"),
    ("torch", "PYSEC-2026-2286", "2.4.1"),
}
  • [x] Step 2: Run schema tests and capture RED

Run:

pytest -p no:cacheprovider tests/test_advisory_baseline.py -q -k 'load_baseline or real_baseline'

Expected: collection or assertions fail because the module, policy, and interfaces do not exist.

  • [x] Step 3: Implement the canonical baseline loader

Use frozen dataclasses and exact constants:

SCHEMA_VERSION = 1
SURFACE_ORDER = (
    "combined-runtime",
    "torch",
    "documentation",
    "atlas-contract",
)

@dataclass(frozen=True, order=True)
class AcceptedAdvisory:
    package: str
    advisory_id: str
    accepted_version: str
    surfaces: tuple[str, ...]

@dataclass(frozen=True)
class Baseline:
    schema_version: int
    audited_surfaces: tuple[str, ...]
    accepted_advisories: tuple[AcceptedAdvisory, ...]

Normalize package names with re.sub(r"[-_.]+", "-", value).lower(). Parse with json.loads, validate exact keys/types/order, and compare the original bytes with json.dumps(document, indent=2, ensure_ascii=False) + "\n". Raise one domain exception such as AdvisoryBaselineError with stable, credential-free messages.

Populate the checked JSON with the 21 identities from Step 1. Every entry currently has surfaces ["combined-runtime", "torch"].

  • [x] Step 4: Run schema tests and capture GREEN

Run the Step 2 command. Expected: all selected tests pass.

  • [x] Step 5: Write observation and comparator RED tests

Add focused tests for:

def test_normalize_accepts_clean_and_vulnerable_pip_audit_payloads(): ...
def test_normalize_collapses_duplicate_raw_records_to_one_identity(): ...
def test_normalize_rejects_malformed_payload_and_dependency_records(): ...
def test_compare_accepts_exact_observations(): ...
def test_compare_fails_on_new_primary_id(): ...
def test_compare_fails_on_package_or_surface_drift(): ...
def test_compare_fails_on_accepted_version_drift_with_advisory_present(): ...
def test_compare_fails_on_accepted_version_drift_with_empty_vulnerability_list(): ...
def test_compare_reports_removed_identity_without_failing(): ...
def test_compare_reports_removed_package_without_failing(): ...
def test_alias_only_and_fix_version_changes_do_not_change_policy_identity(): ...

The Comparison object must expose deterministic errors and notices tuples; success is not errors. Removed identities belong only in notices.

  • [x] Step 6: Run comparator tests and capture RED

Run:

pytest -p no:cacheprovider tests/test_advisory_baseline.py -q -k 'normalize or compare or alias'

Expected: failures identify the missing observation/comparison behavior.

  • [x] Step 7: Implement observation normalization and comparison

Track both resolved versions and observed identity/surface membership. Version drift is detected from the full dependency list, not only vulnerable rows, so a package upgraded to a clean but unreviewed version still fails.

Use deterministic diagnostic categories:

new advisory: <package> <version> <id> on <surface>
accepted version drift: <package> expected <old>; observed <new>
surface drift: <package> <version> <id> expected [...]; observed [...]
reconcile removed advisory: <package> <version> <id>
  • [x] Step 8: Write audit-runner and CLI RED tests

Use a fake subprocess runner to assert:

def test_audit_commands_cover_exact_four_surfaces_and_flags(): ...
def test_documentation_audit_uses_disable_pip(): ...
def test_audit_exit_zero_and_one_are_completed_observations(): ...
def test_audit_other_exit_missing_output_and_malformed_json_fail_closed(): ...
def test_cli_returns_nonzero_for_comparison_errors_and_zero_for_notices(): ...
def test_cli_has_no_accept_or_in_place_write_option(): ...
def test_cli_diagnostics_do_not_disclose_index_credentials_or_temp_paths(): ...
  • [x] Step 9: Run audit-runner tests and capture RED

Run:

pytest -p no:cacheprovider tests/test_advisory_baseline.py -q -k 'audit or cli'

Expected: failures identify missing commands, exit handling, or CLI behavior.

  • [x] Step 10: Implement exact audit surfaces and CLI

Each command must begin with sys.executable -m pip_audit, write JSON to a unique temporary directory, retain the Issue #59 strict flags, and never include --ignore-vuln. Delete temporary outputs through TemporaryDirectory cleanup after parsing.

  • [x] Step 11: Add focused tool manifest and Make contract with RED first

Add tests asserting:

assert Path("vulnerability-audit-requirements.txt").read_text() == "pip-audit==2.10.0\n"
assert "audit-advisories:" in makefile
assert "$(PYTHON) -m scripts.advisory_baseline" in makefile

Run the selected tests before editing the manifest/Makefile, then add the exact pin and target.

  • [x] Step 12: Run Task 1 GREEN gates

Run:

pytest -p no:cacheprovider tests/test_advisory_baseline.py tests/test_makefile_contract.py -q
ruff check scripts/advisory_baseline.py tests/test_advisory_baseline.py tests/test_makefile_contract.py
python -m scripts.advisory_baseline --help
git diff --check
  • [x] Step 13: Commit Task 1
git add security/accepted-advisories.json vulnerability-audit-requirements.txt \
  scripts/advisory_baseline.py tests/test_advisory_baseline.py \
  Makefile tests/test_makefile_contract.py
git commit -m "security: add accepted advisory baseline"

12.18.3 Task 2: Offline policy-to-ledger parity

Files:

  • Modify: scripts/verify_repo.py
  • Modify: tests/test_verify_repo.py

Interfaces:

  • Consumes Task 1 load_baseline and canonical identity/surface semantics.
  • Produces D10 findings under D10.dependency_advisory_baseline.

  • Retains existing D10.dependency_ledger_count, Atlas gitlink, and workflow action-pin behavior.

  • [x] Step 1: Write JSON-policy integrity RED tests

Create temporary-repository tests for missing policy, malformed JSON, unsupported schema, unknown keys, duplicate or unsorted identities, wrong package/version/ID, and non-canonical bytes. Assert all produce D10.dependency_advisory_baseline and that the current repository produces none.

  • [x] Step 2: Run focused verifier tests and capture RED
pytest -p no:cacheprovider tests/test_verify_repo.py -q \
  -k 'dependency_advisory_baseline or dependency_ledger_counts_match_current_doc'
  • [x] Step 3: Add offline policy loading to D10

Import Task 1's public loader rather than duplicating schema logic. Convert loader failures to structured D10 findings; do not propagate exceptions and do not execute any network audit.

  • [x] Step 4: Write unique Markdown identity parity RED tests

Mutate the current accepted table one dimension at a time:

def test_docs_d10_flags_baseline_advisory_id_drift(tmp_path): ...
def test_docs_d10_flags_baseline_package_drift(tmp_path): ...
def test_docs_d10_flags_baseline_accepted_version_drift(tmp_path): ...
def test_docs_d10_flags_baseline_surface_drift(tmp_path): ...
def test_docs_d10_collapses_duplicate_raw_rows_for_identity_parity(tmp_path): ...
def test_docs_d10_excludes_historical_rows_from_baseline_parity(tmp_path): ...

Map Markdown Combined runtime; Torch to ("combined-runtime", "torch"). Parse the audited version and primary advisory ID from every current advisory row. The two duplicated raw IDs must collapse to one unique identity before comparing with JSON.

  • [x] Step 5: Run parity tests and capture RED

Run the Step 2 selection. Expected: new mutation tests fail because D10 does not yet compare unique identities.

  • [x] Step 6: Implement exact two-way parity

Report identities missing from JSON and identities missing from Markdown separately. A current Markdown row with an unknown surface label is malformed rather than silently normalized. Keep all existing raw-count reconciliation intact.

  • [x] Step 7: Mutation-audit the fail-closed boundary

Temporarily mutate each of JSON ID, package, version, surface, duplicate key, ordering, and serialization in isolated copies. Confirm each mutation produces at least one D10 error and restore the original bytes after every probe.

  • [x] Step 8: Run Task 2 GREEN gates
pytest -p no:cacheprovider tests/test_verify_repo.py -q \
  -k 'dependency_ledger or dependency_advisory_baseline'
pytest -p no:cacheprovider tests/test_verify_repo.py -q
make verify
ruff check scripts/verify_repo.py tests/test_verify_repo.py
git diff --check
  • [x] Step 9: Commit Task 2
git add scripts/verify_repo.py tests/test_verify_repo.py
git commit -m "test: enforce advisory policy ledger parity"

12.18.4 Task 3: Isolated required CI gate

Files:

  • Modify: .github/workflows/ci.yml
  • Modify: .github/workflows/docs.yml
  • Modify: tests/test_verify_repo.py

Interfaces:

  • Consumes exact make audit-advisories and vulnerability-audit-requirements.txt from Task 1.
  • Produces unconditional job/context dependency-audit.
  • Produces explicit positive workflow-contract node ID executed by verify-repo.

  • Controller updates GitHub ruleset only after a live PR establishes the context.

  • [x] Step 1: Write exact job-contract RED test

Assert job key/name dependency-audit, ubuntu-24.04, Python 3.11, timeout 20, no job-level if, needs, services, container, env, defaults, or continue-on-error, and exact steps:

Checkout
Set up Python 3.11
Install vulnerability audit tool
Compare dependency advisories with accepted baseline

Checkout must disable persisted credentials and omit submodules. Setup-python must cache pip against:

vulnerability-audit-requirements.txt
requirements.txt
torch-core-requirements.txt
torch-requirements.txt
docs-requirements.txt
atlas-contract-requirements.txt

Install command is exact python -m pip install -r vulnerability-audit-requirements.txt; comparison is exact make audit-advisories.

  • [x] Step 2: Run job-contract test and capture RED
pytest -p no:cacheprovider tests/test_verify_repo.py -q \
  -k 'dependency_audit_job_contract'
  • [x] Step 3: Add the minimal CI job

Add it alongside existing unconditional jobs without modifying their semantics or workflow triggers.

  • [x] Step 4: Write mutation and self-contract RED tests

Reject independently:

  • deleted/renamed job or positive test;
  • omitted/reordered/extra steps;
  • condition, failure mask, alternate shell, job container, or services;
  • removed audited manifest from the cache key;
  • unpinned or alternate audit tool install;
  • altered Make target, --ignore-vuln, or direct inline audit commands;
  • Docker/Compose, Atlas lifecycle, JupyterHub, Ollama, ComfyUI, or localhost commands; and

  • removal of the positive node ID from verify-repo's first explicit pytest invocation.

  • [x] Step 5: Run mutation tests and capture RED

Run the Step 2 selection plus the existing repository-workflow self-contract tests.

  • [x] Step 6: Complete the workflow contract

Add exact tests/test_verify_repo.py::test_ci_dependency_audit_job_contract to the explicit node-ID command before the broad selector. Add dependency_audit keywords to the supplemental selector without relying on that selector as the only guard.

  • [x] Step 7: Add docs-workflow path coverage with RED first

Require the docs workflow to watch:

security/accepted-advisories.json
vulnerability-audit-requirements.txt
scripts/advisory_baseline.py
tests/test_advisory_baseline.py

Extend the existing workflow path regression, observe RED, then update .github/workflows/docs.yml.

  • [x] Step 8: Run Task 3 GREEN gates
pytest -p no:cacheprovider tests/test_verify_repo.py -q \
  -k 'dependency_audit or documentation_workflow'
pytest -p no:cacheprovider tests/test_verify_repo.py -q
python -c "from pathlib import Path; import yaml; yaml.safe_load(Path('.github/workflows/ci.yml').read_text()); yaml.safe_load(Path('.github/workflows/docs.yml').read_text())"
ruff check tests/test_verify_repo.py
git diff --check
  • [x] Step 9: Commit Task 3
git add .github/workflows/ci.yml .github/workflows/docs.yml tests/test_verify_repo.py
git commit -m "ci: require dependency advisory audit"

12.18.5 Task 4: Canonical documentation and projection

Files:

  • Modify: README.md
  • Modify: CONTRIBUTING.md
  • Modify: SECURITY.md
  • Modify: CHANGELOG.md
  • Modify: docs/dependency-contracts.md
  • Modify: docs/conventions.md
  • Modify: docs/architecture.md
  • Modify: docs/maintenance/overnight-2026-07-04.md
  • Modify: docs/manifest.yaml
  • Modify: tests/test_check_docs.py
  • Modify: tests/test_build_docs.py
  • Modify: tests/test_wiki.py

Interfaces:

  • Documents Task 1's policy, Task 2's parity boundary, and Task 3's required context.
  • Produces manifest entry 12.18 immediately after design entry 12.17.

  • Produces identical canonical enforcement/removal language on repository, site, and wiki surfaces.

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

Require the canonical docs to state:

  • security/accepted-advisories.json is the policy artifact;
  • make audit-advisories runs all four surfaces without suppression;
  • new primary IDs and accepted-version drift fail;
  • disappeared IDs are reconciliation evidence, not proof of remediation;
  • JSON and current Markdown rows change together through review;
  • dependency-audit is required alongside the existing two contexts;
  • Issue #63 still owns full locks; and
  • Atlas/service startup is not involved.

Require manifest entries 12.17 and 12.18 to be consecutive and require the same marker language in generated site dependency-contracts.md and wiki 6-1-Dependency-ledger.md (use the actual transform filename returned by the existing test helpers).

  • [x] Step 2: Run documentation tests and capture RED
pytest -p no:cacheprovider tests/test_check_docs.py tests/test_build_docs.py tests/test_wiki.py -q \
  -k 'advisory or manifest or dependency'
  • [x] Step 3: Update canonical documents

Make these exact semantic changes:

  • README repository tree adds security/ with the accepted-advisory policy.
  • CONTRIBUTING workflow adds the local command and required CI signal.
  • SECURITY §13.6 replaces the stale no-gate claim with the implemented boundary.
  • Dependency contracts §6.1.1.4 documents comparison/removal; §6.1.12 no longer defers advisory comparison.
  • Conventions adds an isolated dependency-audit subsection and pre-PR command.
  • Architecture identifies the isolated security signal without changing an architecture diagram unless the canonical prose test proves one is required.
  • OM-047 becomes Fixed with Issue #60's policy/job/ruleset evidence boundary.
  • CHANGELOG adds the new gate and removes the Issue #59 entry's stale future-Issue-#60 sentence.

  • Historical Issue #59 design/plan records remain unchanged.

  • [x] Step 4: Add manifest entry 12.18

- id: issue-60-advisory-baseline-implementation
  number: "12.18"
  title: Issue 60 machine-readable advisory baseline implementation plan
  source: docs/superpowers/plans/2026-08-13-issue-60-advisory-baseline-implementation-plan.md
  • [x] Step 5: Run documentation GREEN gates and inspect projections
make docs-check
make docs-wiki
pytest -p no:cacheprovider tests/test_check_docs.py tests/test_build_docs.py tests/test_wiki.py -q
python scripts/verify_repo.py --check docs --fast
ruff check tests/test_check_docs.py tests/test_build_docs.py tests/test_wiki.py
git diff --check

Inspect generated site/wiki dependency-contract pages and navigation, then leave generated trees ignored and uncommitted.

  • [x] Step 6: Commit Task 4
git add README.md CONTRIBUTING.md SECURITY.md CHANGELOG.md \
  docs/dependency-contracts.md docs/conventions.md docs/architecture.md \
  docs/maintenance/overnight-2026-07-04.md docs/manifest.yaml \
  docs/superpowers/plans/2026-08-13-issue-60-advisory-baseline-implementation-plan.md \
  tests/test_check_docs.py tests/test_build_docs.py tests/test_wiki.py
git commit -m "docs: document required advisory baseline gate"

12.18.6 Task 5: Broad verification, required context, GitFlow, and cleanup

Files:

  • No planned production-file changes; review fixes must be separate, focused commits.
  • Update ignored report/ledger files under .superpowers/sdd/ only for execution bookkeeping.

Interfaces:

  • Consumes all prior tasks.

  • Produces reviewed feature branch, live dependency-audit context, updated GitHub ruleset, feature-to-develop PR, develop-to-main PR, main-to-develop sync PR when ancestry requires it, and clean local/remote state.

  • [x] Step 1: Run the real live comparator

Use the exact current manifests and fixed PyPI service:

make audit-advisories

Expected: exit 0, all four surfaces observed, no new identity/version/surface error. Reconciliation notices are acceptable only if their corresponding ledger update is reviewed before merge.

  • [x] Step 2: Run full local verification

In a canonical NNx wheel environment:

make test
make verify
make lint
make docs-check
make docs-wiki
python -c "from pathlib import Path; import yaml; yaml.safe_load(Path('.github/workflows/ci.yml').read_text()); yaml.safe_load(Path('.github/workflows/docs.yml').read_text())"
git diff --check
git status --short

Record exact pass/skip counts and warnings. An editable NNx run is supplemental only.

  • [x] Step 3: Perform task and broad review

Use the subagent-driven workflow's per-task review packages and final whole-branch review. Fix every Critical or Important finding in one focused final-review fix wave, rerun affected tests, and obtain a clean re-review before publishing.

  • [x] Step 4: Push and open feature-to-develop PR

Push codex/issue-60-advisory-baseline, open a ready PR to develop, and include Issue #60 scope, tests, no-service statement, baseline semantics, and rollback boundary.

  • [x] Step 5: Establish and require the live context

Wait for the feature PR's dependency-audit job to appear and pass. Read ruleset 18620095, then update only required_status_checks so the exact contexts are:

pytest-repository
atlas-consumer-policy
dependency-audit

Preserve strict_required_status_checks_policy, do_not_enforce_on_create, pull-request settings, deletion protection, non-fast-forward protection, enforcement, conditions, and bypass actors. Read the ruleset back and compare every preserved field.

  • [ ] Step 6: Merge feature to develop

Require all applicable checks green, merge by PR, verify the remote feature branch is deleted, and fast-forward local develop.

  • [ ] Step 7: Promote develop to main

Open a separate ready develop to main PR that closes #60. Require dependency-audit, pytest-repository, atlas-consumer-policy, Tier-A, docs, verifier, NNx, Docker build, Pages/wiki, and all other applicable checks green before merging.

  • [ ] Step 8: Synchronize main back into develop

If the release merge makes main not an ancestor of develop, open a content-neutral main to develop PR. Require the same applicable checks, merge it, then prove:

git diff --exit-code origin/main origin/develop
git merge-base --is-ancestor origin/main origin/develop
  • [ ] Step 9: Close bookkeeping and clean state

Check every Issue #60 acceptance box, add PR/main/develop/test/ruleset evidence, mark its project item Done, check #60 in parent #53, and leave #53 open. Fetch/prune, remove merged feature branches, confirm one worktree, no open/superseded PRs, a clean tracked tree, and no scoped service containers.

  • [ ] Step 10: Continue the goal

Move Issue #61 to In Progress, create and push its feature branch from synchronized develop, then begin its design/research cycle without reopening completed Issue #60 work.

12.18.7 Portability Amendment: Selector-Free PyG Audit Projection

Linux PyG wheel resolution can produce local +pt24cpu versions that are valid runtime artifacts but unsupported by strict PyPI pip-audit resolution. Keep torch-requirements.txt unchanged for runtime installation. Split audit input into resolver-safe torch-audit-requirements.txt (core include plus torch_geometric) and pre-resolved pyg-extension-audit-requirements.txt (the four compiled extension pins). Each logical combined-runtime/Torch surface runs strict normal resolution then strict --disable-pip --no-deps supplement resolution and merges the observations. Before network, canonical semantic partition validation reconstructs the exact runtime manifest after removing its selector; missing, extra, changed, duplicate, ambiguous, option, continuation, or non-UTF-8 input fails closed. Cache and docs workflow path contracts include both manifests.