docs: add AGENTS.md describing how we build

[sam-goodwin]

Adds a root AGENTS.md: what the repo is, and how we build in it.

## How we build

Tests drive the work, and real infrastructure runs from the first test. Each change goes through the same loop:

1. Write a failing test at the layer the change affects (see Testing).
2. Make it pass against local workerd with `ALCHEMY_DEV=1`, which is fast.
3. Run the same suites against real Cloudflare without `ALCHEMY_DEV`.
4. Open a PR. CI deploys a preview, runs every suite and records a demo video.
5. Merge, then pick up the next change.

Testing layers, each with a snippet from the real suites:

Layer	File	Command
Story	packages/frontend/src/story.test.ts	pnpm --filter @cafe/frontend test
Scene	packages/frontend/src/scene.test.ts	pnpm --filter @cafe/frontend test
API	packages/backend/test/api.test.ts	pnpm --filter @cafe/backend test
e2e	packages/frontend/test/chat.test.ts	pnpm --filter @cafe/frontend test:e2e

Parallel worktrees, each with its own stages so they don't deploy over each other:

git worktree add ../cafe-<topic> -b feat/<topic> origin/main
export ALCHEMY_STAGE=dev_<topic>          # alchemy dev / deploy
export ALCHEMY_TEST_STAGE=test_<topic>    # Test.make

Also covers demos (pnpm demo, or let the PR record one) and conventions: one file per responsibility in domain folders, named Effect.fn, effect/Config instead of process.env, a browser-safe @cafe/api, Conventional Commit PR titles.

🤖 Generated with Claude Code