No way to assert on tracing output in tests

• Status: Todo
• Kind: Chore
• Authors: jp
• Date: 2026-09-04
• Label: domain=tooling
• Label: type=task

Nothing in the workspace can assert that a tracing event fired. Grepping tracing_subscriber|with_default|logs_contain across jp_llm and jp_test returns no test helper, so a log line is untestable and a deleted one is invisible.

That is a gap wherever a log line is the behaviour. PR #1083 added a warn! in jp_llm::provider::openai_compat::parse_chunk that fires when a stream chunk is discarded — the whole point of the change, since the alternative is diagnosing a silent drop from a blank retry loop. Its tests pin parse_chunk's return value and the error field's round-trip, both of which stay green if the warn! is deleted.

test-log is already a dev-dependency in jp_llm, so the subscriber plumbing is half present. What is missing is a capture layer plus an assertion helper, somewhere reusable — jp_test is the natural home given it already carries mock and macros.

The decision the fix forces

Whether the helper asserts on rendered text or on structured fields:

• Rendered text reads naturally in a test and catches wording changes, but breaks on every rewording and cannot distinguish two events that render alike.
• Level plus fields is stabler and matches how the events are actually written, but a test then says nothing about whether the message is intelligible.

The parse_chunk case wants the level and the provider field, not the sentence. Worth settling before writing the helper, since it decides whether captured events are stored as strings or as field maps.

Also needs deciding: whether capture is per-test (a scoped subscriber, safe under cargo test's thread-per-test) or global. Per-test is the only one that works with parallel execution, and it constrains the API to something the test holds rather than a free function.