feat: request, stage, and delivery metrics with a local Collector example (#279 phase 3)

Problem statement / motivation

Parent: #279 (phase 3). Depends on phases 1 #338 and 2 #339.

Operational metrics that are correct without trace sampling, with explicit timing definitions, bounded dimensions enforced by tests, a backend-neutral local Collector setup, and measured overhead with telemetry disabled vs enabled.

Proposed solution

WIP

Acceptance criteria (maps to #279 phase 3)

• The Collector's debug exporter shows every instrument above with service.name and only allow-listed attribute keys/values.
• Counters and agentic.execution.active finalize exactly once for completed, upstream-failure, disconnected, and shutdown-cancelled requests, over HTTP and WebSocket.
• N requests on one WebSocket connection → N execution.count increments and one connections.active increment.
• Metrics are recorded with OTEL_TRACES_SAMPLER_ARG=0; gen_ai.client.token.usage is absent when upstream omits usage — nothing fabricated.
• The three first-* timings are distinct instruments; a cassette test proves first_upstream_data ≤ first_client_event ≤ first_text for a text response.
• Cardinality test fails the build on any attribute outside the allow-list.
• docker compose -f deploy/otel/docker-compose.yaml up + scripts/tests/otel-smoke.sh pass on a clean machine.
• Overhead table (streaming and non-streaming, disabled vs enabled) in the docs

Non-goals

Dashboards/alert rules; exponential histograms (Collector-side); a Prometheus endpoint in the gateway (OTLP push only); per-tool-name / per-request-model / per-conversation / per-URL labels; OTLP log export.

Start by reading the phase 3 acceptance criteria alongside parent #279 and dependencies #338 and #339. Run docker compose -f deploy/otel/docker-compose.yaml up and scripts/tests/otel-smoke.sh, then map the HTTP and WebSocket metric paths and the required tests. Done means every listed criterion passes, including bounded attributes, timing ordering, sampling behavior, smoke tests, and the documented overhead table.