fix(stack): ship local stack service logs to Analytics through Vector

[avallete]

Summary

The experimental stack's Vector recipe launched Vector with its bundled upstream demo config whenever no configPath was given, which is always the case from the CLI. That config emits a fake syslog line every second to the console and leaves Vector's API disabled, so the recipe's /health readiness never passed in docker and nothing reached Analytics.

The recipe now writes a generated pipeline to <root>/<instanceId>/vector/vector.json in prepare:

• Docker / Podman: a docker_logs source filtered to the stack's com.supabase.stack label, routed by a new com.supabase.service container label to per-service remaps ported from the legacy vector.yaml template (Kong omitted; the stack has none), and HTTP sinks to ${LOGFLARE_URL}/api/logs?source_name=… authenticated with ${LOGFLARE_PRIVATE_ACCESS_TOKEN}. Vector interpolates both from the env vars the recipe already set.
• Native: an API-only pipeline. Native service output is only held in stack memory, so there is no source Vector can read; this is an explicit limitation.

Supporting changes:

• ContainerSpec gains a service label (set by process recipes and the database) and an opt-in engineApi flag. The flag mounts the engine socket read-only at /var/run/docker.sock, or sets DOCKER_HOST to host.docker.internal for TCP or named-pipe endpoints. The endpoint comes from docker context inspect (which honors DOCKER_HOST) or podman info, and the Docker Desktop/Colima socket mapping mirrors the legacy CLI.
• Podman without an active API socket fails Vector's launch with an actionable message instead of a bind-mount error.
• experimental stack start SIDE_EFFECTS updated.

A new Vector.integration.test.ts runs the database, Analytics and Vector in docker and asserts that a database RAISE LOG marker lands in Logflare's postgres.logs source with LOG severity.

Reviewer notes

• Vector still receives engine access when a custom configPath is supplied; nothing in the CLI passes one today.
• TCP engine endpoints are assumed to be a local, plaintext daemon, as in the legacy CLI.

🤖 Generated with Claude Code

[avallete]

Superseded by #6893, which persists the stack's service logs to files and ships them to Analytics directly from the stack host (no Vector in the path), catching up after Analytics has slept.