Forward telemetry in a separate thread

[alexmojaki]

logfire.experimental.forwarding now forwards OTLP payloads asynchronously.

The forward_export_request() and logfire_proxy() helpers now enqueue accepted payloads locally and return immediately instead of waiting for Logfire. This prevents Logfire outages or slow responses from blocking application request handling.

Key user-visible changes:

• Accepted requests return local OTLP success, not confirmation that Logfire received the payload.
• Queue-full or post-shutdown requests return OTLP partial success.
• Each Logfire config manages one queue per destination Logfire URL (usually only one).
• The queues have hard limits of 1000 items and 64mb.
• Forwarding now uses all configured Logfire write tokens, not just the first.
• Only application/x-protobuf and application/json OTLP payloads are accepted, so that the response can match.
• Only content-type, user-agent, and content-encoding from the original request are forwarded, whereas previously all headers were forwarded except for a specific blacklist.
• body can no longer be None.
• Uses the same disk-backed retrying mechanism as normal exports instead of only trying once.

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[cubic-dev-ai]

2 issues found

Confidence score: 3/5

• There is a concrete medium-risk behavior in logfire/_internal/config.py: the shutdown path may lazy-initialize the variable provider, which can trigger remote variable machinery during teardown instead of cleanly shutting down.
• A secondary timing issue in logfire/_internal/config.py uses wall-clock time for timeout budgeting; clock changes could skew remaining time and cause premature or delayed timeout handling.
• Given one high-confidence 6/10 issue affecting shutdown behavior, this carries some regression risk and should be addressed before relying on this as a low-risk merge.
• Pay close attention to logfire/_internal/config.py - shutdown/provider invocation and timeout budget calculations are the key risk points.

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>