feat(ocb): PR-A2 — switch tracecore entrypoint to OCB-generated main

[trilamsr]

Summary

RFC-0013 PR-A2 (sequencing gate for PR-B2 / PR-F / PR-I): retire the hand-wired ./cmd/tracecore entry point and adopt the OpenTelemetry Collector Builder (OCB) output at ./_build/tracecore as the canonical binary. After this lands, all receivers register through OCB's generated otelcol.Factories instead of the bespoke cmd/tracecore/components.go.

Big diff (-3,869 / +388 across 53 files) because PR-A2 is the load-bearing pivot point; PR-B2 / PR-F / PR-I that follow can land surgically.

CLI surface change — `tracecore` now uses the upstream OCB-generated CLI:
- `tracecore collect --config=…` → `tracecore --config=…`
  (collect was default; OCB main runs the collector by default)
- `tracecore receivers list` → `tracecore components`
  (now shows receivers + processors + exporters + extensions + connectors)
- `tracecore debug dump` → removed (no OCB equivalent; use `tracecore
  components` + the live config when filing issues)
- `tracecore failure-inject {nccl-hang,pod-evict}` → use the standalone
  `tools/failure-inject` binary (already ships xid, nccl-hang,
  pod-evict, cpu-steal subcommands)
- `--log.format=text` / `--shutdown.drain-budget=…` / `--version-short`
  → removed; OCB upstream uses `--feature-gates` + `--set` flags

Chart-shape changes — operator-visible:
- Default pipeline flips from clockreceiver→stdoutexporter (in-tree,
  not registered by OCB) to hostmetrics→debug (both upstream
  OCB-bundled). Fresh install on a no-GPU cluster boots and emits
  load-average metrics immediately, same as before.
- `telemetry.listen` + `telemetry.paths.{metrics,healthz,readyz}` →
  `telemetry.metricsListen` + `telemetry.healthListen` +
  `telemetry.healthPath`. The legacy single-listener block is gone
  because upstream `service.telemetry` and `healthcheckextension` are
  two separate processes. Probes hit `:13133/` (healthcheckextension
  default) instead of `:8888/healthz`. Prometheus scrape port (8888)
  is unchanged.
- Self-telemetry metric names rename `tracecore_*` → `otelcol_*`
  (upstream vocabulary). Dashboards on any `tracecore_receiver_*`,
  `tracecore_exporter_*`, `tracecore_queue_*`, `tracecore_component_*`,
  or `tracecore_build_info` must be rewritten — see
  `docs/migration/v0.1-to-v0.2.md` for the exact map.

Breaking changes — orphan components (PR-A2 → PR-J/K bridge)

The OCB-assembled binary registers only the components in
builder-config.yaml: 6 receivers, 4 exporters, 3 extensions,
4 processors. The chart's per-component toggles for the legacy in-tree
set survive this PR so the values shape doesn't break for operators
that pin them, but enabling any of the following in chart values
will cause the pod to fail at startup with an "unknown factory" error
until PR-B2 / PR-J / PR-K rewire them:

Component	Kind	Replacement (planned)
clockreceiver	receiver	hostmetrics (PR-E shipped; now default)
containerstdout	receiver	filelogreceiver + container stanza + file_storage extension (PR-J)
dcgm	receiver	dcgm-exporter DaemonSet + prometheusreceiver (PR-J)
k8sevents	receiver	k8sobjectsreceiver + OTTL k8s.event.hint transform (PR-J)
kernelevents	receiver	journaldreceiver + filelogreceiver (kmsg) + OTTL Xid transform (PR-J)
nccl_fr	receiver	In-repo Go submodule via OCB gomod: (PR-B2 + PR-I)
pyspy	receiver	Deferred until OTel Profiles GA
stdoutexporter	exporter	debug (OCB-bundled; now default)
otlphttp (in-tree clone)	exporter	otlphttpexporter (OCB-bundled; same otlphttp name in chart values, same field shape — endpoint, compression, headers, tls.*, timeout, retry_on_failure, sending_queue; pass-through render so any upstream field works)

To verify what's actually registered in the binary you're running:
./_build/tracecore components.

The chart's NOTES.txt surfaces a WARNING when an operator enables
any of these, and the chart-render CI workflow now runs tracecore validate against the default + one-receiver-on fixtures so a chart
edit that emits a non-OCB key trips CI before reaching helm install.

What landed

Deletions (3,032 LOC across 22 source + 7 test files)

• cmd/tracecore/ entire tree: main.go, collect.go, validate.go, debug.go, receivers.go, signals.go, failure_inject.go, openflags_{linux,other}.go, receiver_variants{,_dcgm_cgo,_dcgm_stub}.go, components.go, + every _test.go (collect_test, debug_test, failure_inject{,_linux}_test, integration_test, integration_telemetry_test, main_test, receivers_test)
• components.yaml + tools/components-gen/{main,main_test}.go — superseded by builder-config.yaml; OCB owns codegen now
• components/receivers/kernelevents/runbook_test.go — depended on the deleted tracecore debug dump subcommand; kernelevents itself is scheduled for deletion in PR-K

Build path swap

File	Change
Makefile	build target now runs OCB (was: legacy go build ./cmd/tracecore); dropped generate, generate-check, run, legacy -ldflags -X version injection; coverage -coverpkg drops ./cmd/...
install/kubernetes/tracecore/Dockerfile	Builds via OCB (make build then copy _build/tracecore)
install/kubernetes/tracecore/templates/daemonset.yaml	args drops the collect subcommand; probes hit the new health port (13133) at healthPath
install/kubernetes/tracecore/templates/_helpers.tpl	renderedConfig emits upstream OTel shape — service.telemetry.metrics.address + extensions.health_check + service.extensions: [health_check] — instead of the legacy single-listener telemetry: top-level block
install/kubernetes/tracecore/values.yaml	Default pipeline flipped to hostmetrics → debug; in-tree-only toggles kept with explicit doc-comments naming PR-J/K as their migration owner
.goreleaser.yaml	Switched to builder: prebuilt against ./_build/{Os}-{Arch}/tracecore; release.yml gains a per-platform pre-build step
.ko.yaml	Builds from inside ./_build/ (the OCB submodule) via KO_CONFIG_PATH=../.ko.yaml; main: .
.github/workflows/ci.yml	package job runs OCB; old build-ocb drift gate replaced by smoke-test-binary job consuming the package artefact
.github/workflows/release.yml	New "Pre-build OCB binaries" step before goreleaser; ko-publish step runs from cd ./_build/ so OCB submodule resolves
.github/workflows/chart.yml	Restored the tracecore validate gate against the default + one-receiver-on chart renders; path triggers swap cmd/tracecore/** → builder-config.yaml
.github/workflows/install-bench.yml	Path triggers swap cmd/tracecore/** → builder-config.yaml
builder-config.yaml	dist.version bumped to 0.1.0-m9-alpha to match Chart.yaml appVersion; chart-appversion-check.sh now reads it as source of truth
scripts/chart-appversion-check.sh	Read dist.version from builder-config.yaml (was: internal/version/version.go)
scripts/smoke.sh	Rewritten for OCB binary — hostmetrics → debug config, expects upstream lifecycle log lines
scripts/validator-recipe.sh	BIN default now ./_build/tracecore
scripts/{doc-check,no-autoupdate-check}.sh	Drop cmd from scan paths

New integration seam

• internal/integration/ocb_scrape_test.go: spawns _build/tracecore against a hostmetrics → debug config, polls the upstream :NNNN/metrics surface, asserts both otelcol_process_uptime and otelcol_receiver_accepted_metric_points are present, then SIGTERMs the subprocess and asserts clean exit. The test skips when _build/tracecore is absent so a fresh git clone + go test ./... stays green; make build is the prereq. This is the regression gate for the chart's operator-facing self-telemetry contract (RFC-0013 §3): if a future upstream OCB release renames either metric, the chart's service.telemetry.metrics.address advertisement breaks downstream dashboards silently — this test fires first.

Integration recipe migration

• docs/integrations/examples/honeycomb.yaml + otel-backend.yaml: clockreceiver → hostmetrics loadscraper (the OCB-supported equivalent per RFC-0013 PR-E). The other two recipes carry pending-rfc-0013-pr-a markers and are still skipped by validator-recipe.sh.

Doc rot fixes

• docs/FAILURE-MODES.md: rerouted 8 entries that referenced deleted in-tree tests to their upstream OCB owners.
• docs/FLAKY-TESTS.md: moved the two in-tree integration flakes to Resolved.
• STYLE.md: rewrote repo-layout + component-registration + CLI + build-release sections around OCB.
• PRINCIPLES.md: dropped concrete-example reference to deleted file.
• install/kubernetes/tracecore/README.md: ko local-build steps now run from inside ./_build/.
• docs/rfcs/0013-distro-first-pivot.md: PR-A2 entry rewritten as landed.
• docs/migration/v0.1-to-v0.2.md: added rows for the self-telemetry metric-name rename (tracecore_* → otelcol_*) and the telemetry.* chart values key rename.

Sequencing constraints honored

• components/receivers/{clockreceiver, containerstdout, dcgm, k8sevents, kernelevents, nccl_fr, pyspy} survive as orphan-but-compiling code until PR-K deletes them along with the chart-fixture migration.
• internal/{pipeline, selftelemetry, telemetry, componentstatus, pipelinebuilder, consumer, fanout, runtime} survive until PR-F (after PR-B1 lifted nccl_fr off internal/selftelemetry in feat(pivot): PR-B1 — port nccl_fr off internal selftel + lifecycle #184).
• tracecore validate gate in chart workflow is restored on the default + one-receiver-on fixtures (was temporarily disabled in the first push of this PR; the chart's renderedConfig template migration was completed in the same PR).

Sequencing gate satisfied

PR-B2 (nccl_fr import swap to upstream OCB types), PR-F (internal/* deletion), and PR-I (Go submodule extraction) are unblocked. The legacy boot path is gone; the OCB-driven boot path is live.

Test plan

• make build → produces ./_build/tracecore binary via OCB
• ./_build/tracecore --version reports 0.1.0-m9-alpha matching Chart.yaml
• ./_build/tracecore components lists 6 receivers + 4 exporters + 3 extensions + 4 processors (the builder-config.yaml inventory)
• ./_build/tracecore validate --config=<rendered chart default> exits 0
• ./_build/tracecore validate --config=<rendered chart one-receiver-on fixture> exits 0
• make smoke passes (hostmetrics → debug, 1.5s window, clean shutdown)
• make check passes
• go test ./internal/integration/... passes (new OCB scrape test; ~1.2s)
• helm lint install/kubernetes/tracecore clean (1 chart, 0 failed; icon advisory only)
• helm template demo install/kubernetes/tracecore --show-only templates/daemonset.yaml renders args: [--config=…] + two ports (telemetry, health) + probes hitting health port at /
• helm template demo install/kubernetes/tracecore --show-only templates/configmap.yaml | yq '.data["config.yaml"]' renders upstream OTel shape: service.telemetry.metrics.address, extensions.health_check, service.extensions: [health_check]
• conftest test --policy install/kubernetes/tracecore/policies/conftest/tracecore.rego /tmp/chart-render.yaml — 51 tests pass
• CI: package job builds amd64 + arm64 OCB binaries
• CI: smoke-test-binary job runs --version + components on the package artefact
• CI: chart workflow lints + templates + validate + yq + conftest pass
• CI: install-bench workflow's kind-cluster install still rolls out (bench values updated to drop the obsolete stdoutexporter reference and align otlphttp.endpoint with the upstream otlphttpexporter schema — pass-through render so all upstream fields work without chart changes)

[trilamsr]

Adversarial cross-cut review found 4 blockers — auto-merge disabled until fixes land. Fixer agent in flight. Summary:

🔴 F1 chart defaults break helm install — values.yaml still defaults clockreceiver (not in OCB binary). validate-CI gate deleted in same PR so this regression has no seam.

🔴 F2 9 in-tree components silently orphaned — release-notes block doesn't disclose breadth. Bench otlphttp config needs schema verification against upstream otlphttpexporter.

🔴 F3 metric-name contract deleted, no replacement test — tracecore_* integration test removed. New OCB binary emits otelcol_* vocabulary. Operators upgrading will see blank dashboards with no migration note.

🔴 F4 chart probes break — /healthz + /readyz not served by OCB binary; healthcheckextension is bundled but not wired in configmap template.

Fixes incoming.