[sehr-m]

Tests that directly call request handlers can update the ViewerContext ContextVar without entering the middleware scope that restores it. The value then survives into later tests in the same worker. After #129261, a full backend run observed UNKNOWN actors in statistical-detector tests that expected organization-scoped SYSTEM actors; the PR's selective run did not include those detector tests. That change was subsequently reverted.

Add an autouse fixture that starts each test with no ViewerContext and restores the incoming value using its ContextVar token in a finally block. A subprocess regression checks isolation after both a successful UNKNOWN-context mutation and a failing USER-context mutation, and verifies restoration of an outer SYSTEM context at session exit. This prepares test isolation without reintroducing the reverted Slack/indexing change or changing production behavior.

Validation:

• Pytest infrastructure and ViewerContext suites: 51 passed.
• Regression fails with the isolation fixture disabled and passes with it enabled.
• Required prek checks and commit/push hooks passed.
• Attempted the original Slack-validation → statistical-detector pair against the reverted Slack code, but local setup stopped on an outdated SCM dependency (missing GitGrantKeyError), before reproducing the original assertions.

Production assessment: normal HTTP handling encloses Slack validation in ViewerContextMiddleware's restoring scope, so the cross-test contamination is not evidence of cross-request production leakage. Direct calls outside a restoring scope can still retain context, and the detector's existing handling of inherited UNKNOWN contexts is unchanged. This PR addresses test isolation only.