[AdaAibaby]

Problem

Fixes #3757.

The reverse proxy in packages/shared/pkg/proxy (shared by client-proxy and the orchestrator sandbox proxy) has no request-outcome metric — only connection gauges (*.proxy.*.connections.open, *.proxy.pool.size). There is no way to tell on a dashboard whether a proxied request failed because the destination could not be resolved (500), the node was unreachable (502), the client went away (499), or the sandbox was not found.

This is the blind spot behind node-churn incidents: during a scale-down, sandboxes on removed nodes surface as Failed to route request to sandbox, and because routing failures (500) and upstream-unreachable failures (502) are indistinguishable at the metric layer, the only way to separate them is scraping logs. #3691 / #3692 split the message; this adds the metric so the split is visible and alertable (a spike in upstream_unreachable by sandbox = stale routing pointing at a drained node).

Changes

Record a request outcome at the points where a request's fate is decided:

• handler() pre-routing terminal branches → client_closed (ctx ended before routing), sandbox_not_found, too_many_connections, upstream_unreachable (owner acquire failed), routing_failed (unexpected resolve error).
• pool ErrorHandler → client_closed (context.Canceled / ECONNRESET / io.EOF) vs upstream_unreachable (dial/transport failure, closed sandbox port); routing_failed when the destination is missing from context.
• ModifyResponse → ok (the upstream answered).

The shared proxy stays telemetry-agnostic: it classifies and calls an OnRequestOutcome(ctx, outcome) callback threaded through reverseproxy.New → pool.New (mirroring the existing ConnectionLimitConfig callbacks). Each service owns its counter + name:

• client_proxy.request.outcome
• orchestrator.proxy.request.outcome

labeled outcome = ok | routing_failed | upstream_unreachable | client_closed | sandbox_not_found | too_many_connections.

Scope note: validation/capacity/resume/token error pages in handler() are deliberately not folded into this counter — they are a separate concern with their own signals (resume-flight, capacity logs). This counter is scoped to the routing/transport outcome, which is the incident-relevant blind spot.

Tests

• outcome_test.go: end-to-end — ok (live backend), upstream_unreachable (dial to a closed port), sandbox_not_found (resolver returns the typed error), asserted via a recording callback.
• Updated all existing reverseproxy.New / pool.New call sites (incl. tests) for the new parameter.
• go test ./pkg/proxy/... and the telemetry desc/unit completeness test pass; go vet + golangci-lint clean on all changed packages; client-proxy and orchestrator build.