docs(agents): rewrite the agent guide with product goals and hard-won invariants

[benvinegar]

Rewrites AGENTS.md (CLAUDE.md already symlinks to it) while the context from the 0.2.x work is fresh. Goals: token-efficient, explains why the project is shaped the way it is, and front-loads the things that actually cost time to learn.

New compared to the old guide:

• Product principles — the feedback loop is the product; one board / one session model is intentional (don't add auth/multi-user); features must work across all three integration tiers; feedback must never be silently lost (the class of bug the session-thread fix addressed).
• Feedback-cursor invariant — agentSeq semantics: what advances it (piggyback, author=user waits), what must never touch it (viewer reads), delivered-once by design.
• SqlStore migration rule — deployed Durable Objects can't be reset; follow the pragma_table_info probe pattern.
• Testing notes that burned us — JsonFileStore returns live objects (capture before update when asserting), the session thread is also a .card (scope e2e selectors), the node:sqlite shim + ambient-types arrangement and why it exists.
• Operational bits — server caches viewer/guide at boot (restart to see changes), release checklist as actually performed for 0.2.0.

Net: 712 words. Everything kept from the old guide survives (entry-point map, runtime-agnostic rule, erasable-TS constraint, sandbox/WebKit warning, validation gate, commit/changelog conventions) — reorganized and tightened.

🤖 Generated with Claude Code