draft: add gateway-backed WebUI chat mode

[AJV20]

Thinking Path

WebUI currently has a direct in-process chat execution path, while messaging surfaces go through Hermes gateway/runtime paths. This draft PR introduces an opt-in gateway-backed browser chat mode so maintainers can review the architectural direction without changing default behavior.

What Changed

• Adds HERMES_WEBUI_CHAT_BACKEND=gateway to route /api/chat/start turns through Hermes Gateway/API Server.
• Adds HERMES_WEBUI_GATEWAY_BASE_URL and HERMES_WEBUI_GATEWAY_API_KEY / API_SERVER_KEY handling for the gateway connection.
• Translates gateway OpenAI-compatible streaming chunks and hermes.tool.progress events back into the existing WebUI SSE events (token, tool, done, stream_end).
• Preserves the existing direct _start_chat_stream_for_session(...) path as the default fallback.
• Documents the env vars in ARCHITECTURE.md and adds static regression coverage for the opt-in branch and streaming mapping.

Why It Matters

This is a reversible architecture slice for making WebUI behave more like Hermes messaging surfaces by delegating execution to the gateway/runtime layer instead of constructing AIAgent directly in WebUI. It is intentionally draft/default-off so maintainers can evaluate config, auth, cancellation, attachment, and runtime-ownership tradeoffs before considering merge readiness.

Verification

Set HERMES_AGENT_PY to the Hermes Agent virtualenv Python for your install, for example ~/.hermes/hermes-agent/venv/bin/python.

• $HERMES_AGENT_PY -m py_compile api/routes.py api/config.py api/streaming.py
• $HERMES_AGENT_PY -m pytest tests/test_gateway_chat_bridge.py tests/test_sprint42.py -q
• Result: 33 passed in 2.20s
• Local manual smoke outside this clean PR worktree: Hermes Gateway /health returned OK and /v1/chat/completions returned the expected gateway-ok response.

Risks / Follow-ups

• Draft only: upstream should review whether this should be an env flag, settings-backed config, or a RuntimeAdapter slice.
• Cancellation currently clears WebUI-side pending state but does not yet forward a stop request to a gateway run id.
• Attachment/image parity is not complete in this spike.
• Approval/clarify/queue/goal controls may need deeper gateway/run API integration before this is ready to merge.
• The state layer affected is the WebUI chat streaming/runtime boundary: WebUI still owns browser presentation/session JSON while gateway owns model/tool execution for opted-in turns.

Model Used

OpenAI Codex CLI via Hermes Agent, with terminal/file tools for implementation, tests, git, and GitHub REST automation.

[AJV20]

Updated the draft branch in 3e51156 after the latest master merge.

What changed:

• Resolved CHANGELOG.md.
• Resolved the api/routes.py conflict by preserving the gateway-backed chat helpers and the new _runtime_adapter_goal_action() helper from current master before /goal handling.

Verification:

• python3 -m py_compile api/routes.py
• /Users/abdiel/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_gateway_chat_bridge.py -q -o addopts= → 3 passed
• GitHub Actions matrix is green for 3.11 / 3.12 / 3.13