Fix reasoning manager response validation

[RasulOs]

Summary

• Harden manager response parsing so final success is bound to the actual <request_accomplished>/<answer> tag and missing/invalid success no longer defaults to true.
• Add shared manager response validation used by both stateful and stateless managers.
• Prevent invalid manager output from being returned after retry exhaustion or retry-call failures.
• Preserve the approved fallback for a valid single plan plus stray final tags by stripping final tags and continuing with the plan.
• Convert non-recoverable manager validation failures into FinalizeEvent(success=False) and make success=None final answers fail closed.
• Clarify manager prompt contracts to require exactly one of plan or final answer.

Tests

• UV_CACHE_DIR=/private/tmp/uv-cache uv sync --python 3.13 --group dev --locked
• UV_CACHE_DIR=/private/tmp/uv-cache uv run ruff check .
• UV_CACHE_DIR=/private/tmp/uv-cache uv run pytest -q -> 239 passed
• UV_CACHE_DIR=/private/tmp/uv-cache uv run python -m compileall -q mobilerun tests
• UV_CACHE_DIR=/private/tmp/uv-cache uv lock --check
• UV_CACHE_DIR=/private/tmp/uv-cache uv run --with black black --check mobilerun/agent/manager/prompts.py mobilerun/agent/manager/manager_agent.py mobilerun/agent/manager/stateless_manager_agent.py mobilerun/agent/droid/droid_agent.py tests/test_manager_response_validation.py

Live regression

• Local Android: emulator-5554; mobilerun ping --device emulator-5554 passed.
• Anthropic OAuth reasoning smoke passed with Android version prompt; no manager validation retry; reported Android 16.
• Gemini OAuth reasoning smoke passed with Android version prompt; no manager validation retry; reported Android 16.
• OpenAI OAuth reasoning smoke was blocked by saved OAuth refresh failure: 401 Unauthorized from https://auth.openai.com/oauth/token. Classified as credential/login blocker, not code regression.
• Cloud command regression passed against ready device 5a4d2cd6-0ba6-49ac-bcdc-d4c709446017: devices --cloud, device ui --cloud, device screenshot --cloud, and device press home --cloud.

Independent reviews

• Claude Opus 4.8 max reviewed the final diff after one temporary 529 Overloaded retry. It found the implementation met the requirements and suggested extra tests for plan+empty/duplicate final tag salvage and run_manager failure wiring; those were added.
• Codex 5.5 xhigh reviewed the implementation. Its empty-plan validation finding was fixed. Its concern about strip-and-continue for plan+final conflicts with the approved plan and Claude accepted that behavior as intentional.