fix(router): NVIDIA inference provider URL path duplication causes 404

[rhuss]

Bug Description

The nvidia inference provider type produces 404 errors from the NVIDIA Build API because of URL path duplication. The provider's default base_url is https://integrate.api.nvidia.com/v1, and the router appends the request path /v1/chat/completions, resulting in https://integrate.api.nvidia.com/v1/v1/chat/completions (double /v1).

The openai provider type works correctly with the same URL pattern, so the path deduplication logic works for OpenAI but not for NVIDIA.

Root Cause

The bug is in crates/openshell-router/src/backend.rs where the upstream URL is constructed by combining the route's endpoint (base URL) with the request path. The path deduplication that strips the /v1 prefix works for the OpenAI provider but does not apply to the NVIDIA provider.

Reproduction Steps

1. Create an NVIDIA provider:

   openshell provider create --name inference --type nvidia --credential "NVIDIA_API_KEY=nvapi-..."

2. Create a sandbox and send a chat completion request through inference.local
3. Observe a 404 page not found response from the NVIDIA API

Verification that the API itself works

Direct curl to the NVIDIA endpoint succeeds:

curl -s https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer nvapi-..." \
  -H "Content-Type: application/json" \
  -d '{"model":"meta/llama-3.3-70b-instruct","messages":[{"role":"user","content":"hello"}]}'

This returns a valid response, confirming the issue is in URL construction, not the API itself.

Expected Behavior

The router should deduplicate the /v1 prefix, producing:
https://integrate.api.nvidia.com/v1/chat/completions

Actual Behavior

The router produces:
https://integrate.api.nvidia.com/v1/v1/chat/completions

This results in a 404 page not found from the NVIDIA API.

Environment

• OpenShell gateway v0.0.98-dev.23 (commit ea690499)
• Deployed on OpenShift in VMs

Agent Diagnostic

• Confirmed direct API access works with curl (valid 200 response from meta/llama-3.3-70b-instruct)
• Traced the routing path: supervisor proxy -> gateway GetInferenceBundle -> router -> NVIDIA API
• Identified the URL construction in crates/openshell-router/src/backend.rs as the source of the duplication
• The OpenAI provider handles the same pattern correctly, indicating the fix should extend the existing deduplication logic to cover the NVIDIA provider type

[rhuss]

Closing: the root cause claim was incorrect.

Code analysis of build_backend_url() in crates/openshell-router/src/backend.rs confirms the /v1 dedup logic handles NVIDIA and OpenAI identically. The 404 experienced during a test deployment was caused by stale inference bundle state during debugging (provider was recreated multiple times), not a URL construction bug.

Apologies for the noise.