fix(core): stop every naming ritual converging on Nova

[RatulMaharaj]

Why

Every agent was naming itself Nova. The naming ritual (ensureIdentity) asks the small model to pick a name with a generic prompt, and LLMs asked to name themselves converge hard on the same modal answer — Nova above all, with Aria, Echo, Sage, and Luna close behind. Temperature isn't a usable lever here since the ritual routes to small reasoning-style models that ignore or reject it, so the fix is prompt-side.

What changed

• The system prompt now names the attractors so the model steers around them: Nova, Aria, Echo, Sage, Luna, Lyra, Aurora.
• Each birth gets three random spark words (drawn from a small evocative wordlist) in the user message, so the search starts somewhere different every time.
• The deterministic fallback pool drops its overlap with the avoid-list (Nova, Echo, Sage → Nyx, Ember, Sol) — determinism kept the pool out of the convergence problem, but an agent told to avoid Nova shouldn't be assigned it.

Existing names are untouched: the ritual runs once and persists, so this only affects future first boots.

Testing

• New test asserts the avoid-list is in the system prompt and that three distinct spark words reach the user message.
• deno task ok — fmt, lint, check, 115 tests pass.

🤖 Generated with Claude Code