Source assertion crosswalk

The initial copied cohort comprised 129 test methods in 21 test modules; these are historical import counts, not the current directory inventory. Each copied module links its exact source revision. At import, detector fixtures retained their original names and assertions except the local adaptations identified below. No failing production source was changed, and no skip, xfail, local waiver or new default allowance was introduced as an import workaround.

The source checkouts were inspected at these clean check-path revisions:

The common source root is the launcher's pinned /workspace. Authored Python roots are tools, tests, repository_checks, nmrnet_inference, and nmrnet_provider. The provider root was added with 003's preparation foundation on 2026-10-18; its nested modules enter the same guard, lint, cache and semantic policy checks. The upstream NMRNetplusplus snapshot is separately owned. Cache and image discovery additionally govern containers. Production-only API scans map to tools, nmrnet_inference and nmrnet_provider; API test-only scans map to both local test suites. This changes inventories, not detector predicates.

Later reviewed Python contracts use the local mechanism in test_python.py, with tests exercising the property in test_python_contracts.py. Annotations leave raw finding predicates and the scan inventory unchanged. A standalone annotation requires a nonblank explanation after --. The ordinary allow-contract scope is one generic expression on a single-line statement or a single-line condition expression in an If header or a For iterable, one outermost generic expression and its generic children in With context expressions or Assign/Return values (including multiline values), the whole value of a Call-valued Expr, or callable signature defaults. Compound bodies, For/With/Assign targets, and callable decorators, type annotations and bodies remain outside those boundaries. The separate allow-lambda-body tag covers only its immediately following lambda default-result expression at matching indentation. The result must itself be a generic candidate; nested lambdas and ENV/CLI inputs remain strict. Signature, sibling and surrounding findings remain independently visible. The For placement covers only a single-line iterable, including when its body starts on that line. Its target, body and else suite remain visible; this dedicated header placement supports neither multiline iterables nor AsyncFor. The existing single-line statement fallback is unchanged. The dedicated For placement uses the same generic-only, operator-input, one-outermost-expression and no-lambda guards. An annotation before a Return covers its value only; sibling expressions, adjacent statements and any enclosing exception-fallback finding remain visible. A Call-valued Expr covers the callee/receiver and arguments of that one call, including multiline calls, under the same outermost-expression, generic-only, operator-input and no-lambda guards. This supports reviewed consumer invocations, not inferred call purity or assertion semantics. Other multiline Expr values, including bare containers, remain unsupported. Existing single-line statement support is unchanged; this narrow syntax extension does not promise formatting invariance for every expression statement. The discarded ENV-pop tag retains its separate, earlier exact-call boundary.

A dedicated annotation immediately before a literal-key Dict entry covers only that entry's value, including multiline values, under the same one-outermost generic-expression and no-lambda guards. All keys in that Dict must be literal constants; any computed key or unpack entry rejects this placement. Exactly one key AST must start on the next line at the comment's indentation. Multiple key ASTs starting on that line and sibling defaults within the value do not qualify. This is an AST key boundary, not a parser for entry punctuation; parentheses can move a separate literal key's AST start onto a later line. That separate entry remains outside the annotated value. Key expressions, sibling entries and enclosing defaults remain independently visible. ENV and CLI inputs retain the ordinary strict guards. This permits reviewed observation sentinels without treating the containing dictionary as one contract; it does not infer that metadata defaults are safe.

A separate detector correction classifies OR-combined comparison predicates as validation results rather than replacement inputs. Mixed value fallbacks and nested defaults, including ENV get defaults and env["PORT"] or 443, remain findings in their own right. This syntactic distinction does not resolve aliases or overloaded comparisons, prove call purity, or establish that validation is correct or fails closed.

A separate applicability correction retains raw OR findings for composed comparison predicates while admitting only their silent-default finding. Comparison leaves may compose through And, Or and Not in any consumer, including assigned values, returns and call arguments. Bare ENV operands, names, calls and literals used as predicate arms do not qualify. Nested defaults and a function-default finding on the same expression remain independently strict. The existing syntactic limits for aliases and overloaded comparisons still apply.

A separate applicability correction retains raw OR findings but admits an OR consumed only as an If branch truth through BoolOp/Not edges when every operand has a recognized ENV read, an ENV-containing comparison, or such evidence through further boolean edges. Opaque calls and other value consumers do not provide that evidence. This covers independent optional fault activations; assignments, returns, call arguments and ORs consumed by comparisons remain strict. Nested defaults and body findings remain independently applicable.

The existing allow-pop-default tag uses the same explanation and placement checks for one discarded ENV pop call with a literal string key, literal None, exactly two positional arguments and no keywords. Only that call's finding is admitted; consumed values, other defaults and adjacent findings stay strict. This declares idempotent revocation, not a replacement input or outcome.

Recognized ENV reads, raw-argument, argparse and args-attribute rules remain strict. A handler in an ordinary try may be reviewed when it contains only pass, declares specific exception names or dotted names (optionally a flat tuple), and neither its header nor protected suite has recognized ENV/CLI access. Only that handler's finding is admitted. Bare, broad or computed types, except* and return fallbacks remain outside this boundary. Exception aliases are not resolved. The checker enforces annotation placement and expression boundaries; human review judges whether the explanation justifies the default. Test-reference parsing and symbol-existence checks are not required. Subsequent fixes, explicit fixture adaptations and qualified applicability decisions are recorded in note 011.

Complete copied Python and operator detector modules

Dapp default allowances are project-specific deployment values: all three local allowance inventories are empty and their exact-inventory assertions remain. The dapp-only COMPOSE_DOWN_CLEANUP silent-success exception was removed. Vote private deployment filenames do not exist here; no local private dotenv file exemption was introduced. The two original secret/example separation tests now exercise the same contract against temporary private/example files. Vote's original empty-get/pop matcher semantics remain unchanged. An empty-get tag alone does not suppress the stronger mapping-default detector. The reviewed pop-tag handling above does not waive Vote's separate production ENV-default rule, which remains strict.

The source dapp scanner tokenizes ENV physical lines. Valid multiline local ENV instructions require continuation joining before tokenization. Both its finding and exact-inventory paths now retain every token and the first physical line; an additional detector fixture proves the three continued defaults are all reported, including continued instructions crossing blank/comment lines without losing later tokens or the first physical line. The source changed-Make-default fixture receives its original allowance only inside patch.dict during that fixture, never in repository scans.

The donor's historical Django migration exemption and its three dedicated fixtures were removed: no nmr_api source enters this repository's inventory. The local production scan, detector, escape and diagnostic privacy checks remain. CommandError detection retains its exact default Django CLI projection assumption; it was not broadened to arbitrary exception classes.

Make and tree assertion mappings

test_make_hygiene.py has 10 methods from the dapp source's 31-method hygiene module. The copied original groups are Python bytecode, first-party symlinks, help inventory equality, hierarchical targets, supported aggregate campaigns, one root Makefile, exact phony inventory, explicit help default, and unique targets. The forbidden-text method's secret patterns remain a separate tenth method. Names map from tests/dapp campaigns to the local test aggregate's exact test/unit test/repository test/integration/provider-core prerequisites; its recipe must remain empty. Generic core unit and contract tests belong to nmr-provider-tools; NMRNet retains installed consumer evidence. Dapp external-network/disposable live scenario disclosures have no counterpart in this local offline aggregate.

The remaining dapp methods are domain-specific: Foundry input permissions, recorded artifact branding exceptions and their negative fixtures, JS and Forge build outputs, staged Solidity formatting, package CI campaigns, ABI closure, Bike token/admin roles and deploy warning, Compose project/UID/update guards, Bike CAM and viewer mock inputs, integration-fuzz descriptor binds, package manifest discovery, Cast RPC secret input, GUI published bind and Anvil port. Branding and license-marker prohibition are not local policy; the upstream snapshot includes its license. Generic symlink discovery is covered by the ported exact-source inventory helper. No JavaScript/TypeScript defaults module was transplanted because no authored JS/TS surface exists here.

test_makefile.py retains 11 of the API source's 13 methods: all nine MakeSourceInspectionTests, its MakeHelpTests method and its recipe-ownership method. _makefile.py is copied as makefile.py. Missing/duplicate define and recipe handling, target/prerequisite distinction, grouped/inline/double-colon rejection, gap preservation and exact direct-owner discovery remain unchanged. The ownership method maps credential publication to packages/image, hostile macro override to hostile SHELL, and destructive default-goal substitution to weights/download; all three original override declarations remain required.

The dapp default-goal assertion now accepts one bare or override assignment to help. Its purpose is explicit safe discovery; API's ownership modifier additionally resists hostile caller assignments. A separate assignment count rejects duplicate authorities even if a second value is not a valid literal. Two added fixtures retain changed/missing-default rejection and prove both accepted spellings and duplicate rejection. The actual GNU Make hostile-caller proof is unchanged; this resolves spelling conflict without weakening ownership. The two API DeploymentDiagnosticsTests assert specific Compose project labels and three API deployment names; there is no corresponding deployment owner.

test_source_inventory.py maps the API source's 10 methods into 13 methods:

source_inventory.py retains every original helper predicate and failure. A host find manifest at /repository-source-inventory/root-entries uses %y\t%f\0 records. Public entries absent from /workspace fail visibly, including future root caches and new files. Checkout control (.git/.agents/ .codex), credentials (.aws) and writable acquisition/runtime state (.work) are explicitly classified without granting their contents. This metadata assertion must not be replaced with the already-mounted entries as its own authority.

Execution requirements and expected discussion failures

Deferred CI disposition (2026-10-14)

The API's unconditional workflow-presence assertion protects its adopted CI lane. This repository explicitly defers CI; when adopted, it will contain minimal critical checks (notes/011). The local presence group now checks an independently reviewed exact workflow inventory, initially empty. A negative fixture proves that an unreviewed workflow cannot expand that scope. All four action-pin, dispatch-default, environment and runner predicates remain unchanged. This is an explicit local applicability decision, not missing-tool forgiveness or a passing claim about CI execution. The original API source and revision remain recorded in test_workflow_posture.py and the assertion crosswalk above.

These modules use Python 3.12 standard library only. Make parser behavior probes require make in the checker image; a missing executable remains a failure. Launcher print probes require Bash and ordinary id utilities. They invoke no engine. The checker requires complete read-only public inputs and the host root entry metadata file. test_conventions.py's exact mount fixture was expanded for those justified grants, with no writable state mount.

Initial port findings included layout-derived roots, default/function fallback code, reversed assertions, absent help/default ownership, absent workflows, missing Make, missing Dockerfile frontend pins/direct FROM digests or nonroot declarations and imported detector code failing a stronger independent ported rule. These are historical port findings, not the current repair inventory; note 011 records completed fixes and applicability decisions. The initial port made no production correction or policy exception.

Remaining current API repository-check modules

This is the complete disposition of the current API modules not copied above. It records missing product owners instead of creating equivalent-looking API deployment, credential or database machinery in NMRNet. Container/scientific families remain separately owned parts of the overall port; a handoff in this table is not a claim that this directory already implements their assertions.

All API tests/repo-checks transparency modules are runtime API behavior: test_configuration_transparency, test_control_status_evidence, test_error_transparency, test_finalize_storage_transparency, test_parser_failure_transparency, test_replay_capacity_transparency, test_retained_capacity_transparency, test_route_memory_transparency, test_service_error_transparency, test_signature_processing_transparency, test_transfer_admission_transparency, test_transfer_filesystem_transparency, test_upload_set_message_transparency, and test_user_error_messages require Django request/services/problem schemas, credential/signature/security state, Upload storage or API capacity owners absent here. Their safe-cause expectation does not authorize copying those product models. test_import_boundaries and test_package_exports assert specific API route/security/operations/composition package owners absent here. test_test_taxonomy hardcodes the API test taxonomy; it is not a universal test organization rule. Local discovery completeness stays with the two local suites, without adopting that different taxonomy.

Foundation repair: runtime proof arguments

The local kernel-grant proof in ../test_container_port_runtime.py retains all eight runtime methods from Magnet repository_checks/test_container_posture.py at 2d4ee50b0359e71900fce9289ea01f46d0f41249: https://github.com/numpde/magnet-deploy/blob/2d4ee50b0359e71900fce9289ea01f46d0f41249/repository_checks/test_container_posture.py

Its unused layout-derived ROOT binding was removed because these checks read kernel and cgroup state, not repository inputs. Assertions place the launcher's expected identity and resource grants before observations, and use explicit msg= diagnostics for identity, resource and retained-build-source failures. The independent expected UID/GID, read-only root, hardened writable tmpfs, loopback-only networking, zero effective capabilities, no-new-privileges, resource limits and missing build-source assertions are unchanged. No detector predicate or default allowance changed.

Foundation repair: explicit Python scan responsibility

The local sources.python_sources helper now requires the keyword-only production_only choice. Root-layout and tuple checks explicitly choose the full authored inventory; runtime-policy, validation-guard and error-transparency checks retain their explicit production selection. Both projections still include every nested future nmrnet_inference module. No inventory root or detector predicate changed. The literal branch that selects the inventory remains a classification, not a substitute for a missing operator input.

Foundation repair: scientific checker repository authority

The dependency, fixture-provenance and scientific-product contract modules now borrow the contained lane's pinned root from quality.sources, rather than inferring a repository from their Python package depth. Their selected input paths are unchanged under the maintained /workspace grant. Expected declaration, lock and solvent facts appear before observations in equality assertions. The original NMRPeak dependency/product and SECS fixture obligations remain unchanged; no lock grammar, scientific selector, provenance predicate or default allowance changed. Fixture-provenance prose now reflects the earned producer and attribution bindings rather than its initial missing-evidence state.