chore(mcp): bump the local stack image to grafana/otel-lgtm 0.33.1 - Pyroscope 2.3.1, the image's healthcheck waits for the stack's ready marker

[using-system]

What

IMAGE moves to grafana/otel-lgtm:0.33.1 (Pyroscope 2.3.1; Grafana 13.2.1, OpenTelemetry Collector 0.160.0, Prometheus 3.14.0, Tempo 3.0.3 and Loki 3.7.7 unchanged), with the six unit-test expectations naming the tag, the env reference's pinned-tag line, and one bullet added to the reference's lifecycle section: LGTM_READY_FILE, the marker run-all.sh writes once every component is up (default /tmp/ready) and that the image's own HEALTHCHECK now waits for before probing the components (grafana/docker-otel-lgtm#1814). The image's compare v0.33.0...v0.33.1 touches, outside examples, CI, tests and tooling files, docker/Dockerfile (the Pyroscope version and the base digest), docker/run-all.sh and docker/healthcheck.sh only, so the env-surface catalog keeps its shape but for that variable.

The healthcheck change touches nothing the server does: odd_stack_up and odd_stack_status judge readiness with their own probes (the four backends through Grafana's datasource proxy, the bounded plugin-wiring wait, the OTLP listener), and docker run / docker inspect neither set nor read the container's health - the reference bullet says so. The skill's verified-versions line needs no change (Grafana stays 13.2.1).

Closes #640

Verified live, on this machine

• MCP server unit suite: 260 passed; ruff 0.16.4 check and format clean.
• MCP server integration suite (bash integration-tests/mcp-server/run.sh) against real Docker on the new image: all seven scripts passed (config surface, stack config, daemon unreachable, env, idempotent up, lifecycle, reset).
• The four signals on the running 0.33.1 container: the demo stack's three services (llmbench-api, llmbench-mcp, llmbench-agent) driven with the stored llmbench-store-load benchmark (2 min, 4,000+ requests), then probed through gcx with the skill's probe_services.py - traces, metrics, logs and profiles present for all three - and one real query per backend (a profile top, a trace ops table, a log count, a counter delta). In the container: /api/health 13.2.1, pyroscope --version 2.3.1, otelcol-contrib --version 0.160.0, Prometheus 3.14.0, Tempo 3.0.3, Loki 3.7.7; /tmp/ready present, the image's health healthy.
• The stack reference check (check_stack_reference.py): 6 references follow the contract. The package validated as CI does: apm install --target claude && apm audit - 127 files scanned, no issue; the tree put back.

Review outcome

One round, no finding. A separate reviewer sub-agent verified with real commands that the image compare changes nothing the catalog describes beyond LGTM_READY_FILE and that the bullet states what run-all.sh and healthcheck.sh do, that stack.py reads neither the Docker health nor the marker and sets no --health-* flag, that no file still names 0.33.0 outside history and generated trees, that the six test hunks replace the tag string only, and that lint and the 260 unit tests pass.

Harness path

Not on the path of /odd-observe, /odd-verify or /odd-status in what a run reads or runs beyond one tag string and one bullet of a reference read at env-configuration time; no measurement owed.

🤖 Generated with Claude Code