feat: artifact factory + codehub init + CI UX fixes

[theagenticguy]

Summary

Ships spec 001 (the laptop artifact factory) end-to-end in a single, self-contained branch.

• Four new skills in plugins/opencodehub/skills/ — codehub-document, codehub-pr-description, codehub-onboarding, codehub-contract-map — all P0 per .erpaval/specs/001-claude-code-artifact-surface/spec.md. The flagship codehub-document runs the codeprobe-pattern four-phase orchestration (Phase 0 precompute → AB parallel → CD parallel → E deterministic assembler) in both single-repo and group mode.
• Six new doc-* subagents — doc-architecture, doc-reference, doc-behavior, doc-analysis, doc-diagrams, doc-cross-repo (group-only). All 8-section scaffold, Sonnet model, tool-allowlisted.
• codehub init — single-command project-scope bootstrap. Copies the plugin into <repo>/.claude/, writes .mcp.json, appends .codehub/ to .gitignore, seeds opencodehub.policy.yaml. Check .claude/ into git and every teammate gets the plugin on clone.
• PostToolUse staleness hook — non-blocking systemMessage after git mutations when graph_hash drifts and .docmeta.json exists. Never auto-regenerates; requires user consent.
• Three ADRs — 0007 (artifact factory), 0008 (codeprobe pattern port), 0009 (output conventions) — record the load-bearing decisions.
• Dogfood fixes (ADR 0010) — three small CLI-UX fixes discovered by running the flow against a private two-repo workspace:
  1. --embeddings now defaults --embeddings-workers to "auto" (single-threaded ONNX on 100k nodes was a 45-min foot-gun).
  2. codehub list gains a HEALTH column that flags dangling registry entries (⚠ missing path) and cleaned indexes (⚠ no graph.duckdb).
  3. Phase 0 of codehub-document now requires a schema preflight — subagents probe information_schema.columns once, cached in .prefetch.md, to prevent Binder Error failures from guessed column names.

Scope decisions locked (not timeline, not revisitable without discussion)

• Self-hosted OSS only. No hosted tier, no managed service, no OpenCodeHub-operated infrastructure.
• No remote / HTTP MCP server. Stdio MCP on the Claude Code plugin only.
• No agent SDK. Agents consume OpenCodeHub via the Claude Code plugin or by shelling out to the codehub CLI.
• No grounding_pack compositor tool, no OpenCodeHub-branded agent, no LLM-based PR review, no IDE plugin / LSP, no fine-tuned models.

Full rationale and forfeits in ADR 0007 + .erpaval/brainstorms/013-synthesis-v2-two-surface-product.md.

What's next after this PR

Spec 002 (CI action surface — OSS GitHub Actions that shell out to the codehub CLI, no HTTP server, no SDK) is at .erpaval/specs/002-agent-grounding-plane/ marked P1. It begins only after this spec gets adoption signal.

Test plan

• pnpm -F @opencodehub/cli build clean
• 18 CLI tests pass (13 setup + 5 init; covers init fresh bootstrap, conflict refusal, idempotent re-run with --force, --skip-mcp, --skip-policy, and every existing setup case)
• codehub init dogfooded against a throwaway tmp repo — copies 28 files, rewrites ${CLAUDE_PLUGIN_ROOT} → ${CLAUDE_PROJECT_DIR}/.claude, writes .mcp.json clean JSON, seeds .gitignore and policy starter
• codehub list dogfooded against the real registry — correctly flagged a pre-existing stale entry
• codehub analyze dogfooded against a real multi-repo workspace (98k nodes, 97 nodes); contracts extracted; group sync worked; empty-case codehub-contract-map artifact handled per AC-5-5
• codehub-document skill invoked from a real Claude Code session against this repo (reviewer validation)
• Claude Code picks up .claude/ project-scope plugin after running codehub init in a clean repo and restarting (reviewer validation)

Commits

• `b61633d` feat(plugin): ship artifact factory (spec 001 P0)
• `7c3d9c9` feat(cli): codehub init bootstraps a repo (project-scope plugin + .mcp.json)
• `7b9e91d` fix: three dogfood findings (embeddings workers, registry health, schema preflight)
• `7d19351` chore: anonymize dogfood references; rename gym Quickwork fixtures; repo-relative paths