[Deep-Sixed]

Work package: the two findings from the v0.3.1 review. Branched from main at 3120337. It keeps contract evidence-v3.2, and the version bump to 0.3.2 follows in its own PR, as #10 did.

commit	finding	fix
cf878d8	P2. Paths and symlink targets were decoded as strict UTF-8. One untracked file named bad\xffname ended run, baseline and attest in a UnicodeDecodeError: exit 1, which means FAIL here, not BLOCKED.	os.fsdecode for git's path bytes and os.fsencode wherever a path or link target goes into the digest. A path that can't be read is now BLOCKED (exit 2), naming the path.
2e78da6	P2. INSTALL.md said "There are no release tags yet" and that a tag would become the thing to pin. Tags exist and can still be moved.	The pin section says to pin a full commit id. A table names each release by its real commits: v0.3.1 → 84418d8, and the runtime it pins, 819b7c2. No tag policy change.

What the tests decided

• No state ID changes. For valid UTF-8 paths, fsdecode/fsencode produce exactly the old bytes. test_valid_utf8_state_ids_are_unchanged_from_the_strict_verifier loads the v0.3.1 verifier and compares state_id and state_paths over a tree with non-ASCII names, a staged rename, symlinks and a nested repository. They are identical.
• drift_paths needs no new encoding. A gate that creates made\xff produces drift that is BLOCKED and named in the record, under both evidence/2 and evidence/3. The record is ASCII ("made\udcff", via json.dumps' default escaping) and loads back unchanged. test_a_record_naming_an_undecodable_path_survives_emit_and_load pins that emit → load_json round-trip, and the bytes survive os.fsencode.
• One silent omission closed. os.walk skipped any directory it could not list. With an unreadable directory inside an untracked nested repository, v0.3.1 reported PASS over a state ID that left that directory out. The walk now raises, and that is BLOCKED.

Regressions

tests/test_path_bytes.py, 11 cases:

• untracked non-UTF-8 filename;
• modified tracked non-UTF-8 filename;
• non-UTF-8 symlink target;
• the same inside a nested repository;
• a gate that creates a non-UTF-8 name, under evidence/2 and evidence/3;
• the emit/load_json round-trip;
• the UTF-8 equality check against v0.3.1;
• an unreadable path (in-process);
• an unreadable file and an unreadable nested directory on disk.

The last two skip as root and run in CI, whose runner is not root.

tests/test_verifier_contract.py: test_install_names_each_release_by_the_commits_it_really_is checks each row of the release table against the repository. The tag must resolve to the named commit, the installer at that commit must pin the named runtime, and that runtime must report the named version.

Verification

On 2e78da6, against a baseline captured on main at 3120337. The baseline gates were green: lint clean, 420 passed / 1 skipped.

• Local: Ruff clean; 430 passed, 3 skipped (420 + 10 new; the 2 permission cases skip as root); drift: false.
• Negative controls:
  • Run as a non-root user against the v0.3.1 verifier, 9 of 11 new tests fail. The two that pass are the equality and round-trip pins, which hold on both by design.
  • All 11 pass as non-root on this branch.
  • The doc test fails on the old INSTALL.md and on a table row naming the wrong runtime pin.
• Unchanged: the invariant block that INV_SHA digests (282f2117…).
• Judgment rungs attested: task, diff, self-review, requirements, architecture, fresh-review.
• CI: push run 36208352454 (attempt 1, head 2e78da6): lint, tests-py311, tests-py314 PASS, imported with import-ci. The pull_request run 36208371594 is also green. I could not read the job logs from this session (the log host is blocked by network policy), so the non-root run of the permission cases above is my evidence for those two.
• Compose (state sha256:0c257351…eb93e2a): 0 blocked required gates, 0 stale, STATE MATCH TRUE, CLEAN CLIMB TRUE, READY FOR HUMAN GATE TRUE.

After this

1. A 0.3.2 release PR: VERSION, pyproject, and a CHANGELOG entry, still evidence-v3.2.
2. A pin-move PR, once that release commit is on main.

Tag protection (a v* ruleset) is a separate administrative step and is not part of this PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_016m6di8xFVdwwq949mJx9U1

---

Generated by Claude Code