test: cover capture invariants on both lanes

[eli-r-ph]

💡 Motivation and Context

Stacked on #1028. posthog-go and posthog-rs hit several capture bugs that tests did not catch, such as a race between enqueue and shutdown. This PR checks the Python stack for each of them and adds tests only where no existing test catches the bug.

Changes

Tests only:

• A before_send that changes nothing leaves the request body unchanged, on both clients and both lanes.
• Enqueue racing shutdown delivers the event and leaves no workers running, on both lanes.
• An AI event at the endpoint's size limit is sent on every send path.
• A failed batch logs one line that names the lane's endpoint and no response text, for all four capture methods.

💚 How did you test it?

• One deliberate bug per check; each fails its new test.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Written with Cursor (Claude Opus) under the direction of the assignee.

🤖 Generated with Claude Code

[eli-r-ph]

@greptileai review

[dustinbyrne]

The test-only hardening looks sound, and existing CI passes for this snapshot. One non-blocking suggestion inline: assert exactly one aggregate loss record in the newly named single-loss-line test.

AI-assisted review.