fix: handle missing usage metadata on premature Anthropic stream termination

[giulio-leone]

Problem

When the Anthropic API stream terminates before sending the message_stop event (e.g., network timeout, connection reset, server error mid-stream), AnthropicModel.stream() crashes with AttributeError.

Crash site (line 412):

usage = event.message.usage  # type: ignore
yield self.format_chunk({"type": "metadata", "usage": usage.model_dump()})

Two failure modes:

1. Empty stream: The event variable is never assigned by the async for loop → UnboundLocalError
2. Premature termination: The last event lacks a .message attribute or .message.usage is None → AttributeError: 'NoneType' object has no attribute 'model_dump'

Root Cause

The code unconditionally accesses event.message.usage after the async iteration loop, assuming the Anthropic stream always completes normally with usage metadata. This assumption fails when:

• Network connection drops mid-stream
• Server times out before message_stop
• API returns an error partway through streaming

Fix

1. Initialize event = None before the loop
2. Use safe attribute access via getattr() chain: getattr(getattr(event, "message", None), "usage", None)
3. When usage data is unavailable, emit zero-usage metadata and log a warning instead of crashing

The agent can still process whatever partial response was received before the stream terminated.

Tests

Added two regression tests:

• test_stream_premature_termination: Stream ends with an event that has no .message attribute
• test_stream_empty_no_events: Completely empty stream (zero events)

All 43 existing tests continue to pass.

Fixes #1868

[giulio-leone]

Friendly ping — handles missing usage metadata on premature Anthropic stream termination, preventing AttributeError crashes when the stream is interrupted.