[lesnik512]

Adds a "Telemetry cost" paragraph to AGENTS.md next to pool and server sizing.

Numbers come from a benchmark of this template in Docker on Linux: Granian with 1 worker and 2 CPUs, Postgres 17, oha at 32 connections, OTel exporting over gRPC to a real collector, Sentry pointed at a dummy endpoint. Medians of 3 x 10 s runs over 2 rounds:

Variant	List decks (rps)	Deck with 20 cards (rps)
Baseline	1452	1711
OTel, every request	990	1167
OTel, parentbased_traceidratio 0.1	1175	1316
Sentry, template config	984	1158
Sentry, Litestar and logging integrations only	1086	1258
OTel and Sentry together	748-813	884-890

The OTel and Sentry together row comes from an earlier run with the same setup.

A py-spy profile under load shows where the time goes. OTel spends it in span creation in the SDK, in the OTLP exporter thread, and in the ASGI middleware (collect_request_attributes, the wrapped send). The middleware runs on unsampled requests too. Sentry, with traces_sample_rate unset, still runs start_transaction, continue_trace and a span per middleware through _create_span_call on every request.

No code change: the SDK already reads OTEL_TRACES_SAMPLER and OTEL_TRACES_SAMPLER_ARG, because lite-bootstrap passes no sampler.