fix(byoa): complete engine registry wiring and default model pins for gemini and qwen

[wg2038]

Summary

When Bring-Your-Own-Agent (BYOA) engines gemini and qwen were added, several downstream mapping and fallback sites were not updated:

1. daemon.ts auth failure hints: authFailureHint() lacked branches for gemini and qwen, falling through to return 'Open Codex on that computer and refresh its login or quota, then wake the agent again.'.
2. daemon.ts triage model resolution: triageModel() lacked branches for gemini and qwen, falling through to <cursor-default> in the LLM cost ledger (llm_calls) rather than using their respective triage models (gemini-2.5-flash-lite and qwen3-coder-flash).
3. registry.ts default model pins: listAgentsForComputer() did not handle CUMORA_DEFAULT_GEMINI_MODEL and CUMORA_DEFAULT_QWEN_MODEL, causing agents without explicit model overrides to receive null instead of the deploy-level default.
4. registry.ts type safety: ENGINE_BINS was declared as Record<string, string>. It is now typed as Record<Exclude<EngineId, 'managed'>, string> (consistent with PAIRABLE) so missing engines are caught at compile time.
5. guard-engine-registry.mjs: Extended to assert authFailureHint, triageModel, listAgentsForComputer, and ENGINE_BINS across all engines in ENGINE_IDS.
6. Documentation & Env Example: Updated docs/COORDINATION.md and .env.example with CUMORA_DEFAULT_PI_MODEL, CUMORA_DEFAULT_GEMINI_MODEL, CUMORA_DEFAULT_QWEN_MODEL, and CUMORA_DEFAULT_ANTIGRAVITY_MODEL.

Testing

• npm run lint: 452 files checked, 0 errors, 0 warnings.
• npm run typecheck & npm run server:typecheck: 0 errors.
• npm run guard:big-brain, guard:llm-tracked, guard:engine-registry: all passed.
• npm test: 1058 passing (added server/src/__tests__/agents-computer-daemon-auth-hint.test.ts and updated agents-computer-engine-detect.test.ts).
• npm run test:integration: 269 passing.

[WhichPaths]

I wrote the gemini and qwen adapters (#99, #113) and the registry guard (#114), so these are my loose ends — thanks for running them down. I checked out the branch and exercised the guard rather than reading it, and the new checks do what they claim:

# deleted the gemini branch from authFailureHint
server/src/agents/computer/daemon.ts → authFailureHint
  'gemini' is missing — add an auth failure hint for 'gemini' in authFailureHint

# deleted the qwen branch from triageModel
server/src/agents/computer/daemon.ts → triageModel
  'qwen' is missing — add triage model resolution for 'qwen' in triageModel

I also checked the failure mode of the four new anchors, since a source-scraping check that silently matches nothing is worse than no check. They fail closed — rewriting return rows.map((r) => { into const mapped = rows.map((r) => { gives:

server/src/agents/computer/registry.ts
  anchor not found: listAgentsForComputer default model fallback — update this guard alongside the refactor

Two things I'd change, both about the fallbacks rather than the wiring:

triageModel's fallback still labels unknown engines as cursor. With the new cursor branch added above it, all nine ids in ENGINE_IDS now return explicitly, so return this.agent.model ?? '<cursor-default>' is reachable only for an engine that is not in ENGINE_IDS — and it writes that engine's cost rows into llm_calls under cursor's name. That is the same mislabelling this PR is fixing, just one engine later. The guard makes the branch mandatory, so nothing legitimate lands there any more; '<unknown-engine-default>' would keep the row honest.

authFailureHint's new default drops the only actionable sentence. The early return above it already covers everything that is not auth-shaped, so the final return is reached only when the detail is auth/quota/billing-shaped and no engine matched. Sending that case to "Check the daemon terminal for details" is strictly less than it had: the operator now learns nothing about the auth failure the regex just confirmed, and the daemon terminal is usually on a machine they are not sitting at — the argument behind #138. Something engine-agnostic but still actionable keeps it, e.g.:

return 'Sign in to that engine\'s CLI on that computer (or fix its API key / quota), then wake the agent again.'

Neither blocks the fix; the wiring itself and the guard coverage look right to me.