fix(observability): wire logging, tracing spans, and FastAPI instrumentation (#64)

[JumpTechCode]

Closes #64.

The OpenTelemetry + structlog observability story was exported but inert at runtime. A TracerProvider existed but nothing created spans, and structured logging never activated.

What was broken

• The lifespan called configure_tracing(settings) but never configure_logging → runtime logging fell back to bare stdlib, structlog never configured.
• span_for_fetcher / span_for_llm had zero call sites → no enrichment or LLM spans ever emitted.
• opentelemetry-instrumentation-fastapi was a declared dependency, but FastAPIInstrumentor was never imported → no request spans.

What changed

• configure_observability(settings) — a single process-level entry point (logging + tracing) in sentinel/observability/__init__.py; the lifespan calls it instead of configure_tracing alone, so the "observability is wired" contract has one place to set up and one to test.
• FastAPIInstrumentor.instrument_app(app) in build_app() — per-app (not the global .instrument()) and idempotent across the many test-fixture builds → a server span per request.
• Real span call sites — span_for_fetcher wraps each fetcher run via a thin _run → _run_guarded split that leaves the tested fetch/timeout/breaker/coerce logic untouched; span_for_llm wraps the whole diagnose_call (end-to-end LLM latency incl. timeout + retry).

Tests

• configure_observability wires both logging and tracing (and runs for real without error).
• build_app() instruments the app (_is_instrumented_by_opentelemetry).
• An in-memory span exporter proves one span per fetcher and one per LLM call.
• The lifespan-wiring test now asserts configure_observability is called — guards the exact gh#64 regression.

505 unit tests pass, 84.84% coverage; ruff + mypy --strict clean. Senior code-review pass: no Critical/Important findings.

🤖 Generated with Claude Code