docs(examples): make gymcoach example runnable end-to-end

[iliassjabali]

Closes #23.

Summary

The gymcoach example shipped a 280-line agent.yaml with no README, no env-var reference, no services to back its memory/health declarations, and two dangling sub-agent refs (./agents/observer.yaml and ./agents/reflector.yaml) that did not exist. GitHub browsers clicking examples/ first had no way to know what this was or how to run it.

This PR goes a bit beyond issue #23 (which only asked for a README) so the example actually delivers the "2 minutes from clone to running code" promise. Files added:

File	Purpose
examples/gymcoach/README.md	Walk-through: prerequisites, one-shell-session quick start, file layout, required/optional env vars, links to migration guide and concept docs
examples/gymcoach/.env.example	Every $env: / $secret: reference in the manifest with comments on their own lines (inline comments break docker compose's env-file parser, and that bit me during verification)
examples/gymcoach/docker-compose.yml	Postgres 16 + Redis 7 with healthchecks, matching the style of packages/control-plane/docker-compose.yml
examples/gymcoach/agents/observer.yaml	Minimal but fully schema-valid sub-agent (parallel monitor for injury risk / scope drift)
examples/gymcoach/agents/reflector.yaml	Minimal but fully schema-valid sub-agent (sequential post-session retrospective)

No changes to agent.yaml itself, no changes to tool_implementations.py (the existing file docstring already flags the stubs clearly, and the README now links to it with the same warning).

Verification

From a clean worktree, I ran the exact README quick start end-to-end:

cd examples/gymcoach
cp .env.example .env          # placeholder GROQ_API_KEY
set -a; . ./.env; set +a
docker compose up -d          # postgres + redis healthy
agentspec validate agent.yaml # ok
agentspec health agent.yaml   # groq reachable 254ms, both sub-agents ok

Total elapsed: 6 seconds from clean state to passing health report. Sub-agents that previously failed with "manifest not found" now pass.

Also verified:

• All 13 relative links in the README resolve to existing files
• Both new sub-agent manifests pass agentspec validate
• docker compose config is clean (no interpolation warnings after moving inline comments off value lines)
• docker compose down -v leaves no leftover volumes
• agentspec audit agent.yaml still runs and returns the same score as before (no regression from the new files)

Out of scope (deliberately)

• Replacing the 10 stub tool implementations with real code. Would require real domain logic.
• Committing a pre-generated LangGraph ./agent/ output tree. agentspec generate is the source of truth for that, and committing its output would defeat the "manifest is the spec" principle in CLAUDE.md.
• Adding a Dockerfile for the agent itself. The example is about the manifest, not containerizing a LangGraph app.

Follow-ups spotted during verification (not in this PR)

• memory.longTerm:postgres and memory.shortTerm:redis report "connection string not resolved ($env:DATABASE_URL)" from the memory health check even though env:DATABASE_URL checks pass in the same run. Looks like a second resolver path in the memory check that does not see process.env. Worth a separate issue.
• service.check.ts explicitly skips loopback addresses with "loopback address is not probed from service checks", which is the right default but should probably be documented somewhere visible.