infra: add QA Docker Compose stack with Postgres + Ollama

[cmeans]

Summary

Self-contained QA testing environment with full embedding support:

• Postgres (pgvector/pg17) — isolated awareness_qa database, UTF-8, named volume
• Ollama (nomic-embed-text) — auto-pulls model on first start, cached in volume
• mcp-awareness — built from local source, port 8421 (alongside production on 8420)

One command up, one command teardown. No more ad-hoc docker run commands for QA.

Usage

# Stand up
docker compose -f docker-compose.qa.yaml up -d --build

# Test
curl -s http://localhost:8421/health | python3 -m json.tool

# Tear down (deletes data)
docker compose -f docker-compose.qa.yaml down -v

QA

Verified

1. • Stack starts cleanly (Postgres healthy → Ollama healthy → app starts)
2. • Health endpoint responds on port 8421
3. • remember creates entries
4. • semantic_search returns results with similarity scores (Ollama embeddings working)
5. • down -v tears down cleanly (containers, volumes, network)

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[cmeans]

LGTM

[cmeans]

LGTM

[cmeans-claude-dev]

Dev Review — PR #229

Clean, self-contained QA stack. No blockers. A few observations for QA to consider:

Observations

1. [Minor] Hardcoded credentials in repo
awareness-qa password is in the compose file. Fine for a local ephemeral stack — just noting it's committed to the repo. No action needed.

2. [Minor] Ollama version pinned to 0.19.0
Good for reproducibility, will go stale over time. Could add a comment noting when to bump, or use a floating tag. Low priority.

3. [Informational] No AWARENESS_DEFAULT_OWNER set
Will default to getpass.getuser() inside the container (likely root). Not a problem, but entries will have owner_id=root rather than a named user. Could set to qa for clarity.

4. [Informational] Auth disabled
AWARENESS_AUTH_REQUIRED not set, so auth is off. Correct for QA ease of use. Just means auth-path testing still needs the production instance or a separate config.

5. [Low risk] Ollama cold-start timing
Healthcheck gives ~360s total (start_period: 60s + 10 retries × 30s). Should cover most cases, but a first-ever pull of nomic-embed-text on a slow connection could exceed this. The cached volume solves repeat starts.

Verdict

Ship it. Nice addition to the QA workflow.

[cmeans]

Thanks for the review. Addressed observation #3 — added AWARENESS_DEFAULT_OWNER=qa so QA entries are identifiable. Other observations noted and accepted as-is (hardcoded creds fine for ephemeral local stack, version pin comment is fair, auth-off is intentional, cold-start timing covered by cached volume).