fix(install-bench): swap to OCB binary for hostmetricsreceiver

[trilamsr]

Root cause

PR #180 (chore(pivot): PR-E unblock — bench heartbeat to hostmetricsreceiver) enabled the hostmetrics receiver in bench/install/tracecore-values.yaml and added it to builder-config.yaml, but the install-bench Dockerfile was still building from ./cmd/tracecore. The generated cmd/tracecore/components.go only registers in-tree receivers — hostmetricsreceiver is upstream OTel-contrib and is only bundled by the OCB-assembled binary at _build/tracecore. The daemonset pod failed config load with unknown component type hostmetrics, kubectl rollout status timed out at 5 m, and bash -e aborted run.sh before any diagnostics fired — so CI showed a bare red with no actionable log.

install-bench has been red on main since 2026-05-31T02:28:40Z and on every PR opened after #180.

Affected PRs (open as of this writing): #186, #187, #188, #189.

Fix

Switch install/kubernetes/tracecore/Dockerfile to build via OCB:

make build-ocb           # generates ./_build/{main.go,go.mod,...} + compiles
cd _build && go build .  # re-link with CGO_ENABLED=0 -trimpath -ldflags "-s -w"

The re-link with our flags guarantees the static binary the distroless base can exec; OCB's intermediate compile uses its own defaults. The final image still uses gcr.io/distroless/static-debian12:nonroot at the same pinned digest.

This is a tactical bridge: PR-A2 (#189) makes _build/tracecore the canonical binary for all builds. Once that lands, the in-tree cmd/tracecore path retires entirely (RFC-0013 PR-F) and this Dockerfile change becomes the new normal across every image, not just install-bench.

Why not the alternatives?

• Revert hostmetrics in bench values → walks back PR chore(pivot): PR-E unblock — bench heartbeat to hostmetricsreceiver #180's pivot intent ("no custom receiver where upstream satisfies"); the legacy clockreceiver is on its way out in PR-K.
• Add hostmetricsreceiver to cmd/tracecore/components.go → diverges the in-tree component list from the OCB-managed one; the whole point of PR-A2 is to delete that divergence.

Bonus: surface root cause on rollout-status failure

bench/install/run.sh had post-deadline diagnostics for the first-data path, but the rollout-status path (the actual failure mode of this regression) just exited via set -e. Added dump_failure_diagnostics() (pod state, kubectl describe, current + previous container logs, rendered config) wired to both failure paths; refactor eliminates the duplicated tracecore-pod spelunking that lived inline. Future regressions surface root cause in the CI log without re-running.

Verification

$ make check
… 0 issues, all modules verified

$ docker build -f install/kubernetes/tracecore/Dockerfile -t tracecore:bench-test .
… exporting to image done

$ docker run --rm tracecore:bench-test components | grep -E "hostmetrics|otlp"
    - name: hostmetrics
      module: github.com/open-telemetry/opentelemetry-collector-contrib/receiver/hostmetricsreceiver v0.110.0
    - name: otlp
      module: go.opentelemetry.io/collector/receiver/otlpreceiver v0.110.0
    - name: otlphttp
      module: go.opentelemetry.io/collector/exporter/otlphttpexporter v0.110.0

End-to-end install-bench (kind cluster + helm install) runs on this PR via the workflow itself.

Cost

Docker build stage adds ~100 s (OCB compile inside Alpine). Bench Docker rebuild only fires on chart/bench/builder-config changes — acceptable.

[CI] install-bench Dockerfile now builds via OpenTelemetry Collector Builder so the bench daemonset can load hostmetricsreceiver; also dumps pod state, logs, and rendered config on rollout-status failure. Unblocks every PR opened after #180.