orchestrator-search-compare:
  GET /research?arm=opencode|mastra&q=...  -> timed SINGLE-orchestrator web-search run
      arm=opencode (default) -> opencode-cli arm; arm=mastra -> streaming-mastra-cli arm
      q=... overrides the default research prompt
  GET /multi-research?q=...                -> BLOCKING multi-agent run (opencode orch
      + >=2 opencode `researcher` workers w/ Parallel MCP); can exceed the HTTP
      timeout on slow opencode runs — prefer the async pair below
  GET /multi-async?q=...                   -> start the multi-agent run on the
      Coordination Workflow (no HTTP wait); returns {instance, runId}
  GET /multi-status?instance=&runId=       -> poll: {status, result, events,
      totalTokens, startedAt, finishedAt}; read until status=completed
  GET /observations?instance=&runId=       -> smoke-captured worker inputs + usage
      emissions (duplicate-billing / registry audit evidence)
  GET /matrix-evidence?instance=&runId=    -> per-unit attempt counts, retryClass,
      commit count, matrix-fault count (Gate 3-5 retry/commit/fault evidence)
  GET /health                              -> {ok:true}
  GET /version                             -> {service, commit, builtAt}
  GET /                                    -> this help