[Deep-Sixed]

Work package: input and state-boundary hardening. Branched from main at 1865ef6. It fixes three findings from the repository review of the merged #2–#5 state, and one observation from the Cerberus qualification (#8). Each was reproduced against the CLI on main first. None of them let bad evidence through: each either crashed, refused good evidence, or said the wrong thing.

Four commits, one per defect, each with a regression that fails against its parent:

commit	defect	fix
14c00a6	Valid JSON that isn't an object ([1], "evidence", 5, null) reached .get and crashed check, compose and import-ci with an AttributeError traceback. #5 fixed this for authority.json only.	load_json refuses anything that isn't an object, as BLOCKED, naming the file and the type.
f8863bb	.verification/ was documented as "excluded whole", but the exclusion listed the files it held before the gates ran. An artifact a gate wrote there was read as drift (BLOCKED) unless .gitignore hid it.	The directory is excluded by prefix. state_id, state_paths and is_dirty share one test that honours it.
a1a45c7	Inside a directory entry, _hash_path followed symlinks. Git reports an untracked nested repository as one dir/ entry, so a link inside it pulled the bytes of a file outside the repository into the state ID.	The top-level rule now applies inside too: an explicit walk that never follows links, rather than rglob, whose symlink behaviour changed between the Python versions CI runs.
2a1e36f	compose reads the baseline beside the first record given. With that record elsewhere, it said "GOVERNING BASELINE MISSING" while the baseline sat beside the others.	Which baseline governs is unchanged. The message names the path it checked and says where one does exist.

Regressions

tests/test_input_boundaries.py, 20 cases:

• Wrong shapes: 16 cases (four shapes × check, compose, and each of import-ci's two payloads), each asserting exit 2, the named reason and no traceback.
• Evidence directory: an artifact written into an un-ignored .verification/ is not drift. A control case, a gate writing outside it, is still drift, naming the path.
• Nested symlinks: outside bytes no longer move the state ID. As counterweights, a retargeted link and an edited real file both still do. It passes on 3.11 and 3.14.
• Baseline message: it names the exact path, and points elsewhere only when a baseline really exists there.

Compatibility contract: stays evidence-v3.2 (decided)

Fixes 2 and 3 change how the state ID is computed in two edge cases: files a gate creates under an un-ignored .verification/, and symlinks inside an untracked nested repository. For every other tree the IDs are the same as before. I've kept evidence-v3.2 and not moved the version. My reading is that these are identity bugs, not a change in what counts as admissible evidence, and a record from before the fix at worst goes stale and fails safe. Decision (maintainer, 2026-09-25): keep evidence-v3.2. No version bump, no installer pin move.

Verification

Climbed on 2a1e36f, state sha256:f092f09e…d2318, against a baseline captured on main 1865ef6:

• Local: 420 passed, 1 skipped (400 + 20 new); Ruff clean; drift: false.
• CI push run 36105948866 (attempt 1, head 2a1e36f): lint, tests-py311, tests-py314 PASS; imported with import-ci.
• Compose: 0 blocked required gates, 0 stale, STATE MATCH TRUE, CLEAN CLIMB TRUE, READY FOR HUMAN GATE TRUE.
• Each new test fails against the parent of the commit that makes it pass.
• The pull_request run 36105970900 is also green.

🤖 Generated with Claude Code

https://claude.ai/code/session_016m6di8xFVdwwq949mJx9U1

---

Generated by Claude Code