test(rekognition): green up PR #1205 — make off-loop read test robust to sys.modules eviction

[groupthinking]

Canonical issue

Remediates the failing test check on #1205 (perf/rekognition-offloop). This PR targets perf/rekognition-offloop directly, so merging it lands the fix on #1205's head and turns its CI green. It carries a single commit (c375307) — no duplication of #1205's perf changes.

Outcome

tests/unit/test_aws_rekognition_provider.py::TestRekognitionDoesNotBlockEventLoop::test_local_image_read_runs_off_the_event_loop_thread fails in the full CI suite (AssertionError: _read_file_bytes was never called) while passing in isolation. This makes the failure deterministic-green by fixing the test's module resolution. The production change in #1205 is correct and untouched.

Root cause — cross-module test pollution, not the diff. tests/unit/test_cloud_ai_integrator.py deletes every youtube_extension.*cloud_ai* key from sys.modules at collection time. The aws_rekognition module key matches, so after that eviction the failing test's late import ... as _rek_mod binds a fresh module object, while the provider instance's _prepare_image_input still resolves _read_file_bytes from the original module's globals. The patch.object landed on the stale module and the read ran unintercepted — correct bytes were returned, but the recording wrapper never fired, so the thread-identity assertion saw an empty observed.

Scope

• Included: patch the method's own __globals__ via patch.dict (the exact namespace _prepare_image_input resolves names from) instead of a re-imported module attribute; drop the now-unused late import.
• Explicitly excluded: no production code change; no other test touched; the sibling test_cloud_ai_integrator.py eviction pattern is left as-is (its del sys.modules[...] is intentional for its own stubbing).

Risk

• Risk level: low
• Failure mode: test-only; __globals__ is a stable CPython attribute of the bound function, immune to sys.modules churn.
• Rollback: revert commit c375307.

Verification

Tied to head c375307:

• Focused test passes in isolation and under the real polluting order (test_cloud_ai_integrator.py first, then the target, one session) — before this change that ordering reproduces the exact CI failure (1 failed); after, 62 passed.
• Full file: 95 passed.
• ruff check on the file: parity with base (the one pre-existing F841 at line 820 is unrelated and unchanged; zero added).
• Required CI — will run on this PR's head.
• Review threads resolved.

Reproduction (local):

# old test, real ordering
pytest tests/unit/test_cloud_ai_integrator.py \
       tests/unit/test_aws_rekognition_provider.py::TestRekognitionDoesNotBlockEventLoop::test_local_image_read_runs_off_the_event_loop_thread
# -> 1 failed: AssertionError: _read_file_bytes was never called
# with this change -> 62 passed

Production evidence

Not applicable — test-only change. The production behaviour (all 14 Rekognition calls off the event loop) is #1205's and is unchanged here.

Agent handoff

• Linked to the PR it remediates (perf(cloud-ai): run AWS Rekognition boto3 calls off the event loop #1205)
• No competing PR implements the same fix
• Required check (test) reproduced red and driven green locally
• Human decision requested: merging this into perf/rekognition-offloop (or cherry-picking c375307 onto it) is left to you — I did not push to perf(cloud-ai): run AWS Rekognition boto3 calls off the event loop #1205's branch directly. perf(cloud-ai): run AWS Rekognition boto3 calls off the event loop #1205 still needs your merge approval to main after CI re-runs green.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
v0-uvai	Ready	Preview, v0	Aug 1, 2026 11:19pm

[coderabbitai]

Important

Review skipped

Auto reviews are limited based on label configuration.

🏷️ Required labels (at least one) (1)

• [‘architecture-gap’, ‘bug’, ‘ci-cd’, ‘ci/cd’, ‘copilot-rabbit’, ‘documentation’, ‘duplicate’, ‘enhancement’, ‘frontend’, ‘github_actions’, ‘good first issue’, ‘help wanted’, ‘high-priority’, ‘invalid’, ‘javascript’, ‘ml-model’, ‘needs-triage’, ‘pipeline-critical’, ‘placeholder-code’, ‘priority:high’, ‘python’, ‘python:uv’, ‘question’, ‘styling’, ‘tests’, ‘v0’]

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository YAML (base), Repository UI (inherited), Organization UI (inherited)

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: afa7e452-8a02-4f3d-ad11-b1d50b434017

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

Agent Completion Truth Gate: BLOCKED

Reasons: invalid_payload

Machine-readable verdict

{
  "details": {
    "invalid_fields": [
      "issue.number",
      "policy.agent_login",
      "policy.run_id"
    ]
  },
  "reasons": [
    "invalid_payload"
  ],
  "verdict": "blocked"
}

Workflow evidence

[groupthinking]

CI triage for this PR

The substantive change here (c375307, test-only) is verified: under the real polluting order — test_cloud_ai_integrator.py then the target test in one session — the previous test reproduces the exact CI failure (AssertionError: _read_file_bytes was never called), and with this change that ordering passes (62 passed); full file 95 passed; ruff at parity with base.

The remaining red checks on this head are all pre-existing, repo-wide gates that are red on the base branch (perf/rekognition-offloop) and on #1205 itself — none is caused by this one-file test diff, and I am deliberately not forcing them green:

• gitleaks (working tree) — the finding is square-access-token matching a PyPI sha256: package hash at uv.lock:5129. A false positive; this diff never touches uv.lock. This is the exact issue fix(ci): stop gitleaks flagging uv.lock PyPI package hashes #1197 / fix(security): scope uv.lock checksum false-positive allowlist (#1141) #1142 are open to fix (scope the checksum allowlist). Not remediable inside a test-only PR without pulling in that unrelated allowlist change.
• agent-completion/truth-gate + Agent completion enforcement — invalid_payload on issue.number / policy.agent_login / policy.run_id, i.e. no filled agent-lock-manifest. This gate is red on perf(cloud-ai): run AWS Rekognition boto3 calls off the event loop #1205 too and is itself being remediated by fix(ci): report Agent Lock gate as neutral when trust policy is unprovisioned #1151 / fix: scope agent gate applicability to real dispatch evidence #1154 ("report the gate neutral when the trust policy is unprovisioned"). I will not fabricate provenance run-id / agent-login values to make a trust gate pass — that would defeat the gate's purpose.

What this PR does deliver: the Python test job is not in this PR's scoped check set (base is perf/rekognition-offloop, diff is test-only), but merging this — or cherry-picking c375307 onto perf/rekognition-offloop — is what turns #1205's test job green. The infra/policy gates above then remain, identically to #1205, and need their own fixes (#1197/#1142 for gitleaks, #1151/#1154 for the agent gate) rather than anything in this change.

Human decision requested: merge this into perf/rekognition-offloop (or cherry-pick the commit), then take #1205 through its own merge approval to main.

---

Generated by Claude Code

[groupthinking]

Contained as a duplicate repair. The verified __globals__-based test fix was transferred onto canonical branch perf/rekognition-offloop / PR #1205 at exact head af04aedb59d5b46871c46ab945536faa58c97d91. This PR is being closed unmerged; its branch and commit c3753073ca0bf6951340edc8f3416487177c2041 are preserved.