fix(scheduled-tasks): reset an explicitly null assistant to lead_agent - #6185

Related: #5288.

Why

PATCHing a scheduled task pinned to a custom agent with {"assistant_id": null} succeeds but silently retains the custom agent. model_dump(exclude_none=True) drops the explicit field before the shared resolver can map it to lead_agent, although creation already supports that default.

What changed

An explicitly supplied null now resets the task to lead_agent. Omitting assistant_id preserves the current pin, even if its custom agent has been deleted. Other null update fields keep their existing ignored behavior. Both cases avoid loading the previous agent.

The change uses Pydantic's model_fields_set and the existing assistant resolver. Tests cover the route and actual SQLite persistence. README and Gateway guidance document the PATCH semantics; the guidance replaces a duplicate nginx-routing sentence to remain within its existing size budget.

Surface area

• Frontend UI
• Backend API — scheduled-task PATCH handling
• Agents / LangGraph
• Sandbox
• Skills
• Dependencies
• Default behavior change — an explicit null PATCH now resets the pin rather than being ignored
• Docs / tests / CI only

Bug fix verification

• Regression path: backend/tests/test_scheduled_task_router_behavior.py.
• Before the handler fix, the explicit-null route and SQLite regressions failed: 2 failed, 49 passed.
• After the fix, both regressions pass. The omitted-field case still preserves a deleted custom agent's pin.

Validation

From backend/:

• Related router, scheduler, dispatch and authorization tests: 124 passed.
• make lint: passed; Ruff formatting check covered 1,744 files.
• make test-blocking-io: 189 passed, both before and after.
• All four offline backend shards, using make test with pytest-split's least_duration strategy: 22,096 passed; optional integrations and live tests skipped. The first shard was rerun after adjusting the guidance note's location/size and passed.
• pytest tests/test_agent_guidance_check.py -q: 13 passed.
• python ../scripts/check_agent_guidance.py --base-ref origin/main --head-ref HEAD: 0 errors; existing soft-size warnings remain.
• git diff --check: passed.

The initial full baseline additionally failed 22 web-fetch cases because this host lacked Node/Readability.js. After adding Node to the test PATH and installing ReadabiliPy's JavaScript dependencies in the local virtual environment, that unchanged test file passed all 48 cases. No dependency or lockfile changes are included in this PR.

AI assistance

Tool(s) used: Codex.

How you used it: Codex inspected the original agent-pinning issue/PR and current code, checked for duplicate contributions, wrote the handler fix, regression tests and documentation, and ran the validation above.

• I've read and understand every line of this change and take responsibility for it — it's not unreviewed AI output.