release: 0.9.2.2 — the gates

[0bserver07]

Cuts 0.9.2.2 — the gates. Rolls the Unreleased accumulator into a named, dated block and bumps both version sites (pyproject METADATA + chimera.__version__, pinned equal by tests/test_version_parity.py).

The name

0.9.2.1 was the verified grader — it added a canary that grades the grader. This batch is the follow-through: graders, published claims, on-disk paths and cwd-relative writes each moved from convention to a gate that can fail. Every one of those moves started by finding a defect the convention had been hiding.

A sub-version, deliberately — the third digit is not a batch counter (playbook 14, rule 0).

Cutting the release found two more defects

A test broken for months that no gate could see. tests/otter/test_server_tls.py raised TypeError on every invocation since e5d4d725 gave serve_http a pidfile_prefix argument its fake never accepted. It sits behind pytest.importorskip("cryptography") — skipped in the full-suite run, skipped in the CI-posture run. Invisible to both standard gates because neither installs that dep.

Fixed here, and the fake now captures and asserts the argument rather than swallowing it with **kwargs: a fake that silently absorbs new parameters stops testing the call contract it exists to test.

uv sync --all-extras is now a required release gate (playbook 14). Two lower postures cannot cover a test that both skip, and a release is the one moment worth paying for the highest posture.

The E2B backend is broken against e2b >= 2.x — filed as #177, disclosed and deliberately not fixed here. Measured against installed 2.30.0: template / api_key / timeout moved off the Sandbox constructor onto Sandbox.create(), while pyproject declares an unbounded e2b>=1.0, so --env e2b raises TypeError at construction. The reconnect branch too — connect is an instance method in 2.x, so Sandbox.connect(sandbox_id, ...) would bind the id to self.

Two blind spots stacked to hide it: the tests inject a fake SDK at the module boundary (right for CI, but a fake never has to match the real signature), and mypy chimera/ is clean in CI precisely because CI installs no extras, leaving the call unchecked. It surfaced only under the new gate.

Not fixed because verifying a two-generation compatibility change needs real E2B infrastructure, and replacing a loud TypeError with an unverified code path would be worse than the present failure. Disclosed in the release notes so an upgrading user meets it in writing rather than at runtime.

How I found the first one

I distrusted my own test count. The README claimed 10735 as of 2026-07-28; I had measured 10,470 on a larger codebase. Tests do not vanish — so the number was wrong, not the suite. Re-measuring under all extras gave 11,017, and the extra tests included one that had never run in CI at all.

Numbers now travel with their posture

posture	result
uv sync --all-extras	11,018 passed / 79 skipped
CI posture (scripts/ci_posture_check.sh)	10,501 passed / 133 skipped
cold mypy chimera/, CI posture	736 files, 0 errors

The 517-test gap is entirely importorskip for deps CI does not install — not a smaller suite, a less-tested one. That gap is exactly where the broken test lived. ruff clean, all 7 trademark scrubs pass, canary sweep 7 pass / 0 BROKEN.

The single local failure in both runs is test_validation_split, documented in CLAUDE.md as env-sensitive locally and green in CI.

Contents

The block covers the grading work merged as b72b7898 (mbpp-plus 99.7% → 84.9%, the length-as-correctness and substring-grading classes, the static AST gates), plus the storage/paths registry with doctor/gc, the claims gate reading git ls-files, interception seams reaching every loop with /resync hot-swap, the experiment toolkit, and the bench-matrix exit-code fix.

Full detail: docs/releases/0.9.2.2.md.

After merge

Tagging v0.9.2.2 triggers the publish workflow to PyPI. That step is held for explicit owner confirmation.

🤖 Generated with Claude Code