refactor: self-contained flow pipeline and pi workflow bridge

[bearmug]

Summary

Make repo-arch a cleaner self-contained CLI flow with explicit subcommands and no flag-based modes.

CLI shape

• repo-arch init — write a starter config
• repo-arch flow run — prepare history/cards/dataset/train plan
• repo-arch flow run full — include embeddings + eval
• repo-arch flow inspect — inspect the latest run
• repo-arch train prepare — export training plan
• repo-arch train cycle — continue the persistent training loop
• repo-arch train resume — resume from the latest checkpoint
• repo-arch train status — inspect training session state
• repo-arch train list — list training sessions
• repo-arch train run — execute one-shot LoRA training

What changed

• removed --full, --run-train, and --run flags from the public flow/train surface
• added src/config.ts and src/flow.ts for a versioned run contract
• run bundles now write stable artifacts under .repo-arch/runs/<run-id>/
• training cycles now persist session state and checkpoints under each flow run
• embedding cache is model-aware
• added a thin pi skill + extension bridge for guided workflow suggestions
• added a Pi install section to the README (npm, git, and pi -e examples)
• fixed the CLI entrypoint so the Homebrew symlinked repo-arch shim works correctly
• added npm publish verification backoff to handle registry propagation delays

Validation

• npm run typecheck
• npm test
• npm run build
• ruby -e "require 'yaml'; YAML.load_file('.github/workflows/publish.yml'); puts 'yaml ok'"

Notes

The CLI stays the source of truth; the pi skill/extension are thin guides over the same contract.