feat: self-host launch readiness (llama.cpp / vLLM, e2e memory check, cleaner public signals)

[vitorcastro78]

What and why

The repo gets almost no traffic, and the traffic it does get sees a stale v0.1.0-beta, an Ollama-first README, and an Issues tab full of bot posts. This PR repositions the launch for people who self-host on llama.cpp / vLLM and makes the first "aha" provable in CI.

• Engines: docker-compose.llamacpp.yml (CPU, pulls a small GGUF) and docker-compose.vllm.yml (GPU) with the flags that matter for tool calling (--jinja, --tool-call-parser).
• Chat aha: scripts/aha-chat.sh / .ps1 sends two requests in one session. The second one carries only the new question, so the answer has to come from gateway memory.
• e2e CI: e2e-llamacpp builds the gateway from the commit, starts llama-server with Qwen2.5-1.5B Q4_K_M, and runs aha-chat.sh. Runs on push to main, nightly, and PRs touching the chat path.
• Public signals:
  • content-cadence writes to the run summary instead of opening issues.
  • sync-mirror only syncs one way (Kortexio to the backup) and gives a clear error when the token is bad.
  • release-please accepts RELEASE_PLEASE_TOKEN.
  • New Conventional Commit PR-title check, issue templates (bug report, engine compatibility report), and a PR template.
• Docs: README leads with "Self-hosted memory gateway for your llama.cpp / vLLM server" and a three-command quickstart. docs/self-host.md gets an Engines section. docs/show-hn.md becomes a launch kit.

The first commit carries Release-As: 0.2.0-beta.

How it was tested

• YAML of the compose overrides, workflows, and issue templates parses.
• bash -n scripts/aha-chat.sh
• e2e-llamacpp on this PR is the real end-to-end check; Docker was not available locally.

Needs an org admin

1. Settings > Actions > General: enable "Allow GitHub Actions to create and approve pull requests", or add a RELEASE_PLEASE_TOKEN secret.
2. Rotate the PEER_SYNC_TOKEN secret (fine-grained PAT, Contents: write on vitorcastro78/ContextMemory). The current one is not recognized by the GitHub API.
3. Consider allowing squash merge only, so the checked PR title is what lands on main.