feat(provider): add Nexforce Router

[fernandovitti]

feat(provider): add Nexforce Router

The value proposition

This PR adds Nexforce Router — Latin America's largest AI inference router — so QM
deployments get one API key that reaches Anthropic, OpenAI, Google, DeepSeek, Moonshot,
Zhipu, and Cloudflare Workers AI, with automatic fallback and per-key cost control.

One OpenAI-compatible endpoint (https://router.nexforce.ai/v1), one NEXFORCE_API_KEY,
and your team can swap between the frontier labs and the best open models without touching a
line of infra. Response headers (X-Nexforce-Requested-Model / X-Nexforce-Served-Model)
make it transparent which model answered each request, and per-key spend caps and fallback
chains keep every conversation predictable.

Why it matters for Latin America

Local billing and compliance are the real unlock. Nexforce handles invoicing, taxes, and
payment infrastructure locally for each country in the region, so teams get compliant,
locally-issued invoices without the FX and foreign-procurement friction of paying an
overseas provider. You keep full visibility into which model served each request and what it
cost — per API key, per conversation.

What changed

• src/model/pi-models.ts — register nexforce in MODEL_PROVIDERS, availability, and serviceability
• src/config.ts — read NEXFORCE_API_KEY into the core config
• src/model/model-credential-store.ts — track nexforce availability alongside the other providers
• src/wiring.ts — surface an environment-provided nexforce key to the credential store
• src/api/routes/admin/model-providers.ts — validate nexforce keys against the router (a 1-token chat probe, since /v1/models is public)
• src/deployment/secret-schema.ts — NEXFORCE_API_KEY secret gate
• src/model/model-catalog.ts, src/api/routes/admin/scope-config.ts — carry the nexforce provider through the model catalog types
• cli/src/config.ts — CLI provider registry, key name, and harness rules
• cli/src/backends/doctor.ts — live nexforce key probe for qm doctor
• cli/src/secrets.ts, cli/src/commands/setup.ts — nexforce key in deployment secrets and the setup wizard
• plugins/admin/public/index.html — nexforce appears in the Admin model-provider onboarding
• cli/templates/deployment/deployment.md — operator-facing provider docs
• Tests updated/added for the new provider across core and CLI

How to configure

Environment: set NEXFORCE_API_KEY (or store it write-only from the Admin page →
Model provider), and it becomes a recognized, validated provider key.

NEXFORCE_API_KEY=nfc_...   # minted at https://marketplace.nexforce.ai/workspace/ai-gateway/ai-gateway-keys

Endpoint: https://router.nexforce.ai/v1 — an OpenAI-compatible Chat Completions API
with streaming, tool calling, and multimodal input. Key validation probes a minimal
chat completion against deepseek/deepseek-v4-flash (the cheapest catalog model).

The nexforce catalog — spanning Anthropic, OpenAI, Google, DeepSeek, Moonshot, Zhipu, and
Cloudflare Workers AI — is discoverable at https://router.nexforce.ai/v1/models (public),
including nexforce/smart-route, which lets the router pick an equivalent model per request.

Verification done

• Live model listing: fetched https://router.nexforce.ai/v1/models — 139 models across
  Anthropic, OpenAI, Google, DeepSeek, Moonshot, Zhipu, Meta, Mistral, MiniMax, Qwen, xAI,
  NVIDIA, and the nexforce/smart-route router model.
• Key auth gating: confirmed /v1/chat/completions rejects a bad key with
  401 invalid_api_key, while the catalog stays public — which is why key validation uses a
  chat probe rather than the models listing.
• Test suite: core typecheck + lint clean; 95 affected core tests and 66 affected CLI
  tests pass; CLI typecheck clean. (A live chat roundtrip with a real nfc_ key is pending a
  test key from the reviewer.)

A note on scope

This is the declarative registration: the provider is recognized end-to-end for key
management, secrets, and validation. The QM runtime streams through the pinned
@earendil-works/pi-ai engine, which does not yet know nexforce, so it intentionally has no
servable base model and MODEL_PROVIDER=nexforce fails fast at config load rather than
silently serving broken turns. Wiring nexforce into the streaming layer is the natural
follow-up once the runtime gains the provider.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}