[SDK Sentinel] Fix sandbox importer module-lock race

[sdk-sentinel-bot]

Caution

This PR contains untrusted AI-generated code. Do not approve or run CI until a maintainer has reviewed the diff. SDK Sentinel verified that GitHub Actions remained approval-gated with zero executable jobs when this PR was opened.

Summary

• CI evidence: failed CI job · workflow run · first known failure

Fix intermittent Python 3.10 sandbox workflow validation failures seen in the [first captured target-branch failure](https://github.com/temporalio/sdk-python/actions/runs/35488741770/job/106019795452). No preceding equivalent success was captured.

Root cause

The sandbox importer routed ordinary imports of already-completed modules back through importlib. On CPython 3.10/3.11, a garbage-collection finalizer import can re-enter _ModuleLock.acquire, remove the outer acquisition's thread bookkeeping, and raise KeyError.

Fix

Return completed modules from the sandbox's sys.modules view for plain absolute imports. Relative, fromlist, missing, and still-initializing modules continue through importlib. Tests cover the fast path, fallback conditions, and module-lock reproduction.

Validation

The causal test failed before and passed after the change. The focused importer suite passed with the runtime-specific test skipped on Python 3.14, both affected transfer tests passed, and fixed repository validation passed. Python 3.10 CI remains required. A separate Windows full-suite timeout is unresolved.

Validation status: ci-validation-required

• Flake confidence: high
• Fix confidence: high
• Value: medium
• Patch scope: standard — 3 changed files; 9.5 KiB
• Local reproduction: unavailable — The exact failure requires CPython 3.10/3.11, which is not installed locally. The causal importlib re-entry behavior was demonstrated on Python 3.14.
• Regression coverage: fail-before-pass-after — The loaded-module fast-path test failed before the source change and passed afterward; the exact module-lock test is present but skips on Python 3.14.
• Unchanged baseline (Poe lint, bridge lint, and editable build) — passed
• Independent candidate (Poe lint, bridge lint, and editable build) — passed

Investigator-run checks

• poe test -s -k workflow_sandbox_importer_repeat_import_skips_import_machinery — passed; 2 attempt(s). The causal check failed before the source change because repeated imports re-entered importlib, then passed after the fast path was added.
• poe test -s -k 'workflow_sandbox_importer_repeat_import_skips_import_machinery or loaded_module_fast_path_requires_completed_modules or workflow_sandbox_importer_loaded_module_import_inside_module_lock' — passed; 1 attempt(s). Two tests passed; the Python 3.10/3.11-specific module-lock test skipped on Python 3.14.
• poe test -s -k workflow_sandbox_importer — passed; 1 attempt(s). Seven tests passed and the runtime-specific regression test skipped.
• poe test -s -k transfer_types_workflow_signal — passed; 1 attempt(s). Both workflow-signal parameterizations passed on Python 3.14.
• .ci-flake-runtime/input/validate.sh — passed; 1 attempt(s). Lint, type checks, docstyle, bridge clippy, and editable bridge build passed.

Required target CI

• Continuous Integration / check-protos (Python 3.10, protobuf 3.x)
• Continuous Integration / build-lint-test (3.10, ubuntu-latest)
• Continuous Integration / build-lint-test (3.10, windows-latest)
• Continuous Integration / build-lint-test (3.10, macos-arm)

Residual risks

• The exact CPython 3.10 module-lock path was not executed locally.
• The fast path intentionally bypasses importlib hooks only for completed ordinary absolute imports; CI should confirm no platform-specific import behavior changes.
• The separate Windows Python 3.14 hard timeout remains unresolved.

Automation provenance

• Investigator workflow
• Base SHA: eb642b14947bd8bcdf8816cffb6a63869803f5c6

If this finding should not be fixed, apply the sdk-sentinel:false-positive label and close the PR. Sentinel will suppress the finding until its affected source changes.

SDK Sentinel recurrence history

• Unique failed CI run attempts: 1
• First occurrence: 2026-09-20 04:25 UTC
• Latest occurrence: 2026-09-20 04:25 UTC
• Recent occurrences:
  • 2026-09-20 04:25 UTC — failed CI job