feat: add Dynamic Workflow task execution policies

[Tsuyoshi Ushio (TsuyoshiUshio)]

Summary

• extend the single Dynamic Workflows design record, FRD 0004, with finalized per-task execution-policy decisions
• add bounded retry, timeout, deterministic backoff, continue-on-error, async workflow handlers, idempotency, authorization, structured status, and observability
• route policy-aware plans through Durable-native retry while preserving compatibility for existing plans and histories
• add a focused workflow-retry-policy sample and end-to-end coverage

Durable client lifecycle fix

Azure Durable Functions Python 2.x closes a durable_client_input client when the decorated function returns. The workflow-enabled SSE endpoint previously returned a lazy stream that retained that already-closing client, so a later tool call failed in get_status_all() with RuntimeError: The client is closed.

The SSE route now receives the host-provided durableClient binding configuration, creates one rich client for the response-stream lifetime, supplies it to all workflow management tools for that turn, and closes it on completion or failure. Non-streaming chat, MCP, declared-trigger, list, and status routes continue to use invocation-scoped injected clients.

Regression coverage includes first and repeated turns, a client closed after the prior turn, concurrent distinct streams, failure cleanup, generated binding metadata, and start/status/list/cancel/terminate routing through the current client.

Validation

• Phase 2 architecture review findings are resolved and the task execution policy design is finalized
• independent review of the Durable streaming ownership fix found no actionable issues
• Ruff and strict MyPy pass
• CI-equivalent suite: 1,238 passed, 57 deselected
• local Functions host started with the preserved DTS azureManaged configuration
• natural-language SSE chat invoked start_workflow; DTS executed two transient retries and reached Completed
• a later same-session SSE turn invoked list_workflows successfully, proving a fresh live client is used across turns

This remains a draft while the complete implementation is reviewed. Follow-up to #163; planning issue: https://github.com/Azure/azure-functions-bucees-planning/issues/1278

[Tsuyoshi Ushio (TsuyoshiUshio)]

Updated through 9b208f3 to use Durable Python 2.x native Activity retry for new workflow histories while preserving the legacy Activity/timer driver for marker-free persisted histories. The migration includes strict sanitized exhaustion decoding, native status schema v4, attempt=None where the preview SDK has no attempt visibility, customer-sample/E2E separation, and preview extension/dependency setup.

The branch is synchronized with main and merge conflicts are resolved. Real Functions-host E2E covers retry-to-success and exhaustion with the original sanitized error_code/kind restored after the final TaskFailedError. Read-only skill operations are configured for automatic execution while skill scripts continue to require approval.

All current checks pass: public build and unit tests on Python 3.13/3.14, E2E on Python 3.13/3.14, documentation, SDL analysis, and CLA. The local canonical gate passed 1,235 tests with Ruff and strict mypy clean.