fix: preserve fallback provider attribution on llm node spans

[chenghao-mou]

After failover, llm_node could report the fallback model with the primary provider. Preserve the serving-provider update from #7137 by recording the node's request identity before nested inference starts.

Addresses AGT-3551: issue.

The full local type check reports existing plugin dependency errors also present on main; the core package passes.

Initial prompt and agent context

Model: GPT-6

can you create draft PR for the fallback issue?

The requested issue is the llm_node provider overwrite found during the preceding telemetry review. The fallback stream records the serving provider, then node completion overwrites it with the original provider. This PR preserves the per-request attribution.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[longcw]

from claude code:

What it misses. On the request where the failover happens, FallbackAdapter.model still returns the primary: _next_instance() marks the primary unavailable only after it fails. FallbackLLMStream._run then writes gen_ai.provider.name and gen_ai.response.model to _caller_span, but not gen_ai.request.model. The node span ends with the primary's request model beside the fallback's provider — the same mismatched pair, reversed. The

[chenghao-mou]

from claude code:

What it misses. On the request where the failover happens, FallbackAdapter.model still returns the primary: _next_instance() marks the primary unavailable only after it fails. FallbackLLMStream._run then writes gen_ai.provider.name and gen_ai.response.model to _caller_span, but not gen_ai.request.model. The node span ends with the primary's request model beside the fallback's provider — the same mismatched pair, reversed. The

Yeah, this is expected from this change:
started:

• gen_ai.request.model: model A
• gen_ai.provider.name: provider A

after fallback:

previously:

• gen_ai.request.model: model A
• gen_ai.response.model: model B
• gen_ai.provider.name: provider A

now:

• gen_ai.request.model: model A
• gen_ai.response.model: model B
• gen_ai.provider.name: provider B

so the fallback (requested A but B responded) is clear.