Fix under-specified sample dependencies and stale test/CI wiring

[brianstrauch]

Summary

Four independent fixes. Three share a root cause: a sample's dependencies are satisfied only because another group happens to install them, so the sample breaks the moment its own group is used in isolation. The repo's single shared resolution hides this.

google-adk never declared mcp. google_adk_agents imports McpToolset, but google.adk.tools.mcp_tool imports it inside a try/except ImportError. A missing mcp therefore doesn't raise — the symbol is simply never exported and the sample's toolsets silently disappear. It works today only because google-genai and strands-agents pull mcp in.

google-genai declared mcp>=1.0.0 while the sample imports mcp.shared.session, which mcp 2.0 removed. Nothing in that group caps mcp, so on its own the declaration selects 2.0.0. Demonstrated with a scratch project holding only that group's dependencies:

"mcp>=1.0.0"      -> mcp 2.0.0    (mcp.shared.session gone)
"mcp>=1.27.1,<2"  -> mcp 1.29.0

Added the llm-stream group. workflow_streams/README.md tells readers to run uv sync --group llm-stream for scenario 5, but the group was never created, so the command fails. That sample's openai import resolves today only via the langfuse/langsmith groups.

The gevent test hasn't run in CI since July 2025. It was gated on matrix.python == '3.12', which stopped matching when a5ee599 moved the matrix to 3.10/3.13 — two matrix bumps ago. The gate was originally != '3.7' (gevent didn't support 3.7); once 3.7 was dropped the guard was obsolete, but it was rewritten in positive form as == '3.12' and relabelled "on latest", which is what let it rot. Dropping the condition restores the original intent and can't go stale again.

Also included

• Lock moved to mcp 1.29.0 and google-adk 2.6.3. Both sat at older versions only because uv lock doesn't upgrade already-locked packages; nothing constrained them.
• tests/activity_sticky_queues/ → tests/worker_specific_task_queues/. The sample was renamed long ago; its test directory named a sample that no longer exists.
• tests/google_genai/mcp_test.py locates its echo MCP server through the installed package instead of Path(__file__).parents[2], which hardcodes the distance from test to sample. A wrong path here makes the test hang rather than fail, since the client waits on a subprocess that never started.
• __init__.py added to the 10 test directories that lacked one. Test filenames collide across samples by design (workflow_test.py, activity_test.py and mcp_test.py each appear in several), and package markers are what keep pytest from treating them as the same module. The ones without a marker worked only because they hadn't collided yet — and the rename above adds exactly such a filename.

Test plan

• poe test: 150 collected, 148 passed, 4 skipped — collected count identical to before
• poe lint: clean
• Each changed group verified in a pruned environment (uv run --exact --no-default-groups --group <g>), which is what surfaced the missing/loose declarations
• The four mcp-sharing samples (google_genai, google_adk_agents, strands_plugin, openai_agents) run together on mcp 1.29.0: 24 passed
• gevent_async/test/run_combined.py run on both ends of the matrix, 3.10 and 3.14

One thing CI will exercise that I couldn't check locally: the gevent step now runs on all 8 matrix jobs including windows-latest, where it hasn't actually executed since July 2025. If it's unhappy there, gating on runner.os != 'Windows' is the obvious fallback.

🤖 Generated with Claude Code