fix: test_tracing_backend.py tests always skip (Telemetry not initialized)

[planetf1]

Problem

All 6 tests in test/telemetry/test_tracing_backend.py always skip with reason Telemetry not initialized, even when OTel is installed and Ollama is running.

Found whilst working on PR #742. Flagging for @ajbozart as the original author of the telemetry tests.

Root Cause

mellea/telemetry/tracing.py sets _tracer_provider at module import time based on the MELLEA_TRACE_BACKEND env var:

_tracer_provider = None
if _OTEL_AVAILABLE and (_TRACE_APPLICATION_ENABLED or _TRACE_BACKEND_ENABLED):
    _tracer_provider = _setup_tracer_provider()

The test's setup_telemetry fixture uses MonkeyPatch.setenv("MELLEA_TRACE_BACKEND", "true"), but by the time the fixture runs, tracing.py is already imported and _tracer_provider is already None. setenv cannot retroactively change a module-level constant.

The span_exporter fixture then checks if provider is None: pytest.skip(...) — which always fires.

Attempted Fix

Patching tracing._backend_tracer and tracing._TRACE_BACKEND_ENABLED directly in the fixture (bypassing the env var) gets 4/6 tests passing. The remaining 2 have deeper design issues:

• test_span_duration_captures_async_operation: calls trace.get_tracer_provider().force_flush() on the global OTel provider, not the test provider — flushes the wrong thing. Span name search for "chat" may also not match.
• test_context_propagation_parent_child: creates its parent span via trace.get_tracer(__name__) (global provider) but backend spans go to the test provider. Parent span never reaches InMemorySpanExporter. Context propagation assertion requires both spans on the same provider.

Suggested Fix

In span_exporter fixture, directly patch the module internals rather than relying on env var:

@pytest.fixture
def span_exporter():
    from mellea.telemetry import tracing

    provider = TracerProvider()
    exporter = InMemorySpanExporter()
    provider.add_span_processor(SimpleSpanProcessor(exporter))
    test_tracer = provider.get_tracer("mellea.backend")

    old_tracer = tracing._backend_tracer
    old_enabled = tracing._TRACE_BACKEND_ENABLED
    tracing._backend_tracer = test_tracer
    tracing._TRACE_BACKEND_ENABLED = True

    yield exporter

    exporter.clear()
    tracing._backend_tracer = old_tracer
    tracing._TRACE_BACKEND_ENABLED = old_enabled

test_context_propagation_parent_child additionally needs to create its parent span using the test tracer (not trace.get_tracer(__name__)) so both spans land on the same provider.

Verification

With the partial fix above, 4/6 tests pass against a live Ollama instance. The 2 remaining failures require the test bodies to be updated.

[ajbozarth]

Added as a sub-issue of the tracing refactor epic