[MaxGhenis]

Follow-ups to the reviews of microcosm#1118 (G1, the Axiom graph adapter, merged 4121f27). Findings 1–6 of review-microcosm-1118/out-053144.md and finding 6 of out-033221.md.

• Submodules. A RuleSpec root containing a gitlink is refused, clean or edited; edits inside a submodule that git status misses no longer pass.
• Inherited git environment. Git runs with the repository-local environment variables removed, and the root must be its repository's top level, so GIT_DIR/GIT_WORK_TREE cannot point the check at a clean decoy or another HEAD.
• Byte-level checkout check. Committed paths, unfiltered bytes and executable modes are checked against the declared commit, which also catches edits hidden by index flags, the stat cache, a clean filter, a filesystem monitor, a nested .git entry or a replace ref. SHA-256 repositories are accepted and checked byte for byte.
• Module path. The resolved module path is recorded with the reference; a link repointed after the reference blocks the next compile and a second reference.
• Kernel hash. It depends only on the set of bound adapter classes, not on binding order or ref names; classes sharing a __qualname__ are ordered by module. Both surviving mutants from the review (sort by ref name; sort by qualname alone) are now killed.
• Docs. Native text and date outputs are documented as numpy arrays; the changelog's checkout claim matches what is checked.

Invariants (Hypothesis property tests)

• A referenced adapter compiles only from bytes that hash to the recorded tree digest (test_property_it_compiles_only_from_the_referenced_bytes).
• A git root is accepted iff its index, worktree and HEAD are exactly the declared commit with no gitlinks or flagged entries, whatever the caller's git environment (test_property_accepted_iff_the_root_holds_exactly_the_commit).
• The kernel implementation hash depends on the set of adapter classes only (test_property_the_hash_depends_only_on_the_bound_classes).
• Differential: computed blob ids equal git hash-object (test_differential_blob_ids_equal_git_hash_object).

Tests (build lane)

test_axiom_graph_adapter.py and test_rules_kernels.py, one file at a time: 183 passed, 4 skipped (the existing real-engine graph tests; no local engine). Test evidence by kind is listed under "Review fix" below. ruff check/format and tools/ci_test_plan.py verify clean. Full report: lane report (hub copy at ~/reviews/nz-hub/jobs/new-g1b/report-cont2.md).

Out of scope: an Axiom engine CI job (plan R12), the household-only-weights contract (G2), frozen graph files, frame/kernels.py.

Shared files and owners

packages/microcosm-frame/src/microcosm/frame/adapters/axiom.py and rules_kernels.py tests only; axiom.py is also edited by draft #826 and #822 (Max's sessions). No uk_runtime/us_runtime, frozen graph file or frame/kernels.py change.

Part of the Microcosm NZ v0 graph plan (Max ruling d962), package G1b. Built by the NZ hub on Subfleet build lanes: Claude Opus 5.5 wrote most of it and a GPT-6.1 Sol lane finished it after a provider error (both in the commit trailer); the hub verified scope before committing.

Review fix (head b06217f)

GPT-6.1 Sol review (Subfleet 20261008-095817) at 0c7e90d: REQUEST_CHANGES for one defect, fixed in b06217f and independently verified (APPROVE_WITH_NITS).

• Unicode-equivalent names. Git on macOS with core.precomposeunicode=true records precomposed names (usually NFC) while the file system lists the NFD name a file was created under; comparing the two path sets as strings refused that clean checkout. Committed paths now pair with files one to one: identical names first, then the single canonically equivalent file the committed path opens, confirmed with Path.samefile. Ambiguous and unmatched names are refused, one directory entry never answers for two committed names, and the byte, executable-mode and directory-digest checks are unchanged. Names that differ only in case are not equivalent.
• Scope of the exact-commit claim. The check reads the commit's paths and blob names from the repository's object database and assumes that database is intact; deliberate tampering with it (a substituted tree object) is out of scope, and the recorded directory digest still pins the bytes present.
• Test evidence by kind. Regression tests, red on main's aafc628f adapter source and green here: the submodule, inherited git environment, module-path, hidden byte/mode and replace-ref checkout tests. Regression tests red on 0c7e90d and green here on macOS: the decomposed-name acceptance, hidden-change and property tests. Positive controls that pass on main by design: test_native_outputs_are_numpy_arrays_text_and_dates_included and test_a_module_link_left_alone_compiles_the_referenced_file. Mutant coverage for the unchanged rules_kernels.py: all three kernel-hash tests fail under the sort-by-reference-name mutant; test_classes_sharing_a_qualname_are_ordered_by_module and the property test also fail under sort-by-qualname-alone.
• Tests at b06217f: test_axiom_graph_adapter.py 162 passed, 5 skipped (4 need a local Axiom engine; 1 needs a normalization-sensitive file system); test_rules_kernels.py 29 passed. Not observed on Linux: the three macOS end-to-end tests are expected to skip in CI, and the distinct-NFC/NFD-files test to run there.

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at b06217f with --merge. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: b06217f. Independent review: in-session Opus 5.5 builder + separate verifier (jobs/wf-fixes-1008-journal.jsonl) returned APPROVE_WITH_NITS at b06217f; nits applied in that head (one-to-one NFC/NFD pairing confirmed by samefile, changelog scoped to an intact object database). Earlier G1b Subfleet review findings resolved. All 10 checks green, MERGEABLE/CLEAN, no CHANGES_REQUESTED.