fix(tracing): stop display sessions feeding back on the dashboard itself

[thiagoesteves]

What

Fixes the intermittent failures in the tracing/version test files (promised in #64) - and the root cause turned out to be a real product bug, not just test timing.

The bug

Display trace sessions send every message to the requesting LiveView, whose handle_info calls ObserverWeb.Common.uuid4() for stream ids (and plenty of Enum while rendering). Tracing any module the dashboard itself uses (ObserverWeb.Common, Enum, ...) therefore feeds back on itself: every rendered trace message generates the next traced call, in a storm that runs until max_messages aborts the session. In tests this raced the assertions (flaky by seed and machine load); for real users it silently kills exactly the kind of session "trace Enum" describes.

Tracer.Server.handle_trace/2 now drops display events originating from the request_pid. Tool sessions (count/duration/call-seq/flame-graph) only render at the end, so they keep the requester's calls.

Test hardening

• Version server tests synchronize on the mocked RPC + :sys.get_state instead of Process.sleep(100); the :warning test now actually asserts the warning status (it previously asserted nothing and said so in a comment).
• Tracing page/tracer tests wait for trace messages and session-state transitions with bounded polling or explicit messages instead of fixed sleeps; Request Function terminate synchronizes on session creation before waiting for the reset (its old 50ms sleep could also leak a running session into the next test).
• Tracer unit tests invoke traced functions from a Task, since the requester's own calls are now (correctly) excluded.

Verification

• 12/12 randomized-seed runs green on the previously flaky files (they failed roughly 1 in 3 before).
• 5/5 randomized full-suite runs green (395 tests).

Risk assessment

• Impact: display sessions no longer report calls made by the dashboard's own LiveView; all other origins unchanged.
• Blast radius: one guard in Tracer.Server.handle_trace/2; everything else is test-only.
• Regression risk: low - the excluded origin was self-observation noise that could abort sessions; coverage 96.0%, credo/sobelow/dialyzer/format clean.
• Rollback plan: revert the commit.

Checklist

• mix test green across repeated randomized runs
• mix coveralls 96.0% (threshold 95%)
• mix credo --strict, mix sobelow, mix dialyzer, mix format --check-formatted clean
• Small focused diff, no leftover debug output

🤖 Generated with Claude Code