Phase 2: Pluggable TaskStore on create_a2a_server (PR-N)

[bokelley]

Summary

adcp.server.create_a2a_server() today instantiates InMemoryTaskStore() directly. Tasks don't survive restarts. Downstream (salesagent) has 2,288 LOC of custom A2A server code largely because this hook doesn't exist — they can't adopt the SDK's A2A path without losing DB-backed task persistence.

Proposed API

from a2a.server.tasks.task_store import TaskStore

def create_a2a_server(
    handler: ADCPHandler,
    *,
    ...,
    task_store: TaskStore | None = None,  # defaults to InMemoryTaskStore()
) -> Starlette: ...

Downstream subclasses a2a-sdk's TaskStore ABC (or implements the protocol) against their own Postgres/Redis/etc. and passes it in.

Reference impl

Ship examples/a2a_db_tasks.py with a SQLAlchemy + asyncpg reference. Not production-quality — just enough to prove the hook works and give the next adopter a starting point.

Acceptance

• task_store kwarg threaded into DefaultRequestHandler(task_store=...).
• Integration test that uses a file-backed TaskStore subclass and proves the cache survives server restart.
• Example file referenced from docs/handler-authoring.md.

Context

Roadmap item PR-N (Phase 2) from .context/sdk-adoption-roadmap.md. Downstream feedback in /Users/brianokelley/Developer/salesagent/.conductor/lima-v16/.context/sdk-server-adoption.md — one of three blockers (with #225, #226) making A2A adoption net-negative today.

Revisit A2A adoption story: when N + O + P have shipped and been used by at least one downstream, the ~13-day salesagent A2A migration drops to ~5-day and net-positive.