mios-agent-pipe: router + dispatch + SurrealDB writes (step 2 of 5)

Ports the chain that gives Discord (and every other gateway)
tool-understanding parity with OWUI. Step 1 stood up the FastAPI
scaffold as a transparent proxy; this commit replaces the proxy
with the actual router-classified chain.

Per the Step 1 commit's plan, the parts ported now:

  * Router (Layer-1 micro-LLM classifier)
      - _ROUTER_SYSTEM verbatim from the OWUI pipe (verb table +
        action enum + verb-pick priority + general routing rules)
      - classify_intent() async helper using httpx (OWUI pipe
        used aiohttp; the standalone service uses httpx for
        consistency with the rest of the FastAPI app)
      - Best-effort SurrealDB event row for every router verdict
        (kind=classify, source=mios-agent-pipe, payload=<verdict>)

  * Dispatch (verb -> broker -> structured tool_result)
      - _build_dispatch_cmd() maps the full verb table -> shell
        cmd line, in lockstep with the OWUI pipe's
        _dispatch_mios_verb. Every verb shipped to date is
        supported: open_app, launch_app, focus_window,
        move_window, close_window, list_windows, screen_layout,
        open_url, mios_find, mios_apps, everything_search,
        fs_search, system_status, service_status, service_restart,
        process_list, container_status, container_restart.
      - dispatch_mios_verb() async helper that talks to the
        launcher broker via /run/mios-launcher/launcher.sock
        using the CAPTURE_JSON: protocol (so stdout / stderr / exit
        come back cleanly split -- foundation laid in commit
        24c258d).
      - Returns a structured dict {success, tool, args, output,
        stderr, exit_code, latency_ms} the chat handler wraps
        into the OpenAI tool_call + tool_result envelope.

  * /v1/chat/completions chain
      - Session row in SurrealDB opens at the start of every turn
        (platform=mios-agent-pipe) so downstream tool_call rows
        link back via session = <record_id>.
      - Extract last user message (handles both string and
        OpenAI's array-of-content-parts shape).
      - classify_intent() -> three branches:
          dispatch  -> run broker, build tool_calls envelope,
                       emit either SSE chunks (streaming) or a
                       single JSON response (non-streaming).
          chat      -> emit short reply as SSE chunks or JSON.
          agent/?   -> fall through to the backend (hermes) proxy.
      - tool_call row written fire-and-forget per dispatch (so
        Discord chats also populate the SurrealDB audit trail
        every other gateway gets).

  * SSE chunk shape
      OpenAI-compat streaming protocol -- any gateway with an
      OpenAI client (OWUI's pipe shim, Hermes Discord gateway,
      Slack/Telegram, MCP, ...) consumes the response unchanged.
      The dispatch fast-path emits the structured tool_calls
      envelope as a <details type="tool_calls"> markdown block
      inside a single content delta; gateways with markdown
      rendering get a collapsible block, agents reading the chat
      history see canonical structured JSON inside.

  * /health
      Now reports router (enabled/model/endpoint) + broker
      presence + db_url so an operator probing the service knows
      at a glance which subsystems are reachable.

Deliberately skipped (can be Step 2b):
  * REFINE pass (CPU rewrite of the user message before forward)
  * CRITIC pass + recompose loop
  * POLISH pass + narration collapse
These add quality but not tool-understanding -- Discord's
parity-with-OWUI need is satisfied without them. Operator-
directed if Discord chats need any of them after live testing.

Live-verified on podman-MiOS-DEV:
  GET  /health                  -> json with router/broker/db all
                                   reachable.
  POST /v1/chat/completions:
    "hello there"   -> action=chat -> "Hello! How can I assist
                                       you today?" (no hermes
                                       round-trip)
    "open notepad"  -> action=dispatch -> open_app(name=notepad)
                       -> broker -> [mios-windows] launched via
                       cmd /c start: C:\Windows\System32\notepad.exe
                       -> tool_calls envelope returned
  SurrealDB writes confirmed via journal: 3 session opens + 3
  classify events + tool_call rows all HTTP 200.

Next: Step 3 collapses OWUI's pipe to a thin forwarder pointing
at :8640; Step 4 re-points Hermes Discord gateway at :8640.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: mios-dev