fix: accept OpenAI-chat shape under /gemini and /openai prefixes (#1)

[prasadus92]

Problem

Reported in #1: an OpenAI-chat client (Hermes, or any client using an openai_chat transport) configured with a /gemini base URL fails with:

HTTP 404: Error code: 404 - {'detail': 'Not Found'}

even though curl against the native Gemini route works fine.

Root cause

OpenAI-style clients build their final request URL by appending a fixed suffix onto base_url:

• /chat/completions for inference
• /models, /v1/models, /v1/models/{id} for model discovery

So a base_url of http://127.0.0.1:8787/gemini makes the client call POST /gemini/chat/completions. But /gemini is the proxy's native generateContent route prefix, which has no chat/completions handler. The OpenAI-compat handler was only mounted at /openai/v1/chat/completions, /v1/chat/completions, and /chat/completions. The proxy log in the issue shows exactly this: every /gemini/chat/completions POST and every /gemini/.../models discovery probe returning 404.

curl worked because it used the native shape (/gemini/v1beta/models/...:generateContent); the OpenAI-chat client uses a different shape against the same prefix.

Fix

Mount the OpenAI-compat chat handler and the model-discovery endpoints under the /gemini and /openai prefixes as well as the bare root. Gemini already routes through Vertex's OpenAI-compat endpoint inside _handle_openai (which dispatches by the request body's model, not the URL prefix), so it does not matter which prefix the client appends to.

Newly accepted shapes:

Method	Paths added
POST	/gemini/chat/completions, /gemini/v1/chat/completions, /openai/chat/completions
GET	/models, /openai/v1/models, /openai/models, /gemini/v1/models, /gemini/models, /api/v1/models, plus /{openai,gemini}/v1/models/{id} probes

These join the existing /{,openai/}v1/chat/completions, /chat/completions, and /v1/models routes, matching the project's existing "accept client URL variants" approach.

Docs

• README Hermes recipe + examples/hermes-config-snippet.yaml previously recommended the broken /gemini + openai_chat pairing. Updated to point the openai_chat Gemini provider at /openai (clearest match for the shape) and note that /gemini and the bare root now work too.
• Added a troubleshooting entry for the 404 {'detail': 'Not Found'} symptom.
• Endpoints table + status list updated.

Tests

• test_gemini_prefix_accepts_openai_chat_shape: regression for Not able to use proxy for gemini-2.5-flash #1. The three new chat paths return 200 and forward to Vertex's OpenAI-compat endpoint with model: google/gemini-2.5-flash.
• test_model_discovery_under_client_prefixes: discovery probes resolve under each client prefix; unknown-model probe still 404s.

Full suite: 22 passed, ruff check + ruff format --check clean.

Closes #1