feat(telemetry): expand pipeline and model observability

[guima-why]

Summary

• add pipeline selection timing boundaries and source-channel attributes for observability reports
• record Alibaba Cloud API action/outcome metadata, first-token latency, and normalized token/cache usage across retries, refusals, and fallbacks
• strengthen A2A telemetry cleanup and verify backup-only session restore with a real Step 4 restart path

Root cause and impact

Several reporting dimensions were either missing or attributed too coarsely: pipeline selection timing did not use the precise user-input boundaries, provider usage could be associated with the wrong fallback request, and API outcomes lacked stable action/request metadata. The backup E2E also retained the primary session, so it did not prove that restore actually used the configured backup.

This change adds request-scoped telemetry ownership, explicit timing and usage fields, and a backup-only A2A recovery test. The real E2E exposed and fixed a pre-dispatch hydration gap where an omitted taskId could be replaced before the restored sidecar became available.

Validation

• make lint
• make test (11106 passed, 2 skipped)
• focused A2A regression suite (338 passed)
• real scenario1-performance-backup E2E with DashScope qwen3.6-plus, real tools, and real Alibaba Cloud calls
• verified all restore checks, server logs, and normal-chat recovery; both E2E-created ROS stacks were deleted successfully