feat: initial alpha scaffold — FastAPI + MCP + pgvector

Implements v0.0.1 of OpenBrain:

- FastAPI REST API (capture, search, recent, stats) with Bearer auth
- MCP server (streamable HTTP at /mcp) exposing all 4 tools
- pgvector schema (vector(1024) for bge-m3)
- asyncpg connection pool with lazy init + graceful close
- Ollama embedding client with fallback (stores thought without vector if Ollama unreachable)
- Woodpecker CI pipeline (lint + kaniko build + push to Gitea registry)
- Portainer/Swarm deployment compose
- Mosaic framework files: AGENTS.md, PRD.md, TASKS.md, scratchpad

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs/PRD.md

@@ -0,0 +1,70 @@
+# OpenBrain — Product Requirements Document
+
+**Version**: 0.0.1
+**Status**: Active
+**Owner**: Jason Woltje
+
+---
+
+## Problem
+
+AI agents and tools have no shared persistent memory. Every session starts from zero.
+Platform memory (Claude, ChatGPT, etc.) is siloed — each tool can't see what the others know.
+This forces constant context re-injection, burns tokens, and prevents compounding knowledge.
+
+## Goal
+
+A self-hosted, agent-readable semantic brain that any AI tool can plug into via MCP.
+One database. Standard protocol. Owned infrastructure. No SaaS middlemen.
+
+## Users
+
+1. **Jason** — primary human user, captures thoughts from any AI tool
+2. **AI agents** — Claude Code, Codex, Claude Desktop, any MCP-compatible client
+3. **Future**: Mosaic Stack integration as the knowledge layer for the agent fleet
+
+## Requirements
+
+### v0.0.1 (Alpha — Current)
+
+| ID | Requirement | Priority |
+|----|-------------|----------|
+| R1 | Capture a thought with content, source, and metadata | Must |
+| R2 | Generate vector embedding via Ollama (bge-m3) | Must |
+| R3 | Semantic search by meaning (cosine similarity) | Must |
+| R4 | List recent thoughts | Must |
+| R5 | Usage stats (total, embedded, by source) | Must |
+| R6 | REST API with Bearer token auth | Must |
+| R7 | MCP server (streamable HTTP) exposing all 4 tools | Must |
+| R8 | Deployable as Portainer/Swarm stack | Must |
+| R9 | CI/CD via Woodpecker (lint + build + push) | Must |
+| R10 | Graceful embedding fallback (store without vector if Ollama down) | Must |
+| R11 | Public repo — zero secrets in code | Must |
+
+### v0.1.0 (Future)
+
+- Thought tagging and tag-based filtering
+- Batch import (ingest jarvis-brain data, Claude memory, etc.)
+- Scheduled re-embedding for thoughts stored without vectors
+- Webhook capture endpoint (ingest from any tool without MCP)
+- Usage dashboard (thoughts/day, source breakdown)
+- Mosaic Stack integration (knowledge module backend)
+
+## Acceptance Criteria (v0.0.1)
+
+1. `POST /v1/thoughts` stores a thought and returns it with embedded=true when Ollama is reachable
+2. `POST /v1/search` with a natural-language query returns semantically relevant results
+3. `GET /v1/thoughts/recent` returns the last N thoughts in reverse chronological order
+4. `GET /v1/stats` returns total count, embedded count, and source breakdown
+5. MCP server at `/mcp` exposes all 4 tools (capture, search, recent, stats)
+6. Claude Code can connect to the MCP server and execute all 4 tools
+7. Portainer stack deploys both brain-db and brain-api successfully
+8. CI pipeline runs on push to main and produces a tagged image
+
+## Out of Scope (v0.0.1)
+
+- User accounts / multi-user
+- Workspace isolation
+- Web UI
+- Rate limiting
+- Mosaic Stack integration