Capture descriptions: support any OpenAI-compatible vision endpoint #4846

Problem

hyperframes capture describes downloaded images only through OpenRouter, Vertex or the Gemini API (captionImagesWithGemini in packages/cli/src/capture/contentExtractor.ts). None of them is reliably reachable from mainland China, and none can point at a self-hosted vision model. Users in that position get the catalog-fallback asset-descriptions.md on every capture, even when they already pay for an OpenAI-compatible vision endpoint or run one locally.

Proposed solution

Bring back the three variables from #1809 on current main: HYPERFRAMES_VISION_BASE_URL, HYPERFRAMES_VISION_API_KEY and HYPERFRAMES_VISION_MODEL.

• All three set: caption through <base>/chat/completions with the OpenAI-style request the OpenRouter branch already sends, ahead of OpenRouter, Vertex and Gemini.
• Some but not all set: warn naming the missing variables (never echoing values), skip captioning, and report the phase as degraded. Never fall back to another provider.
• OpenRouter shares the same request helper and keeps its URL, auth header and error text unchanged.
• postExtractionPhase counts a complete custom endpoint as vision credentials, and the no-credentials hint mentions it.

I have this ready and will open a PR that links here.

Alternatives considered

A named preset per regional provider. Every one of them speaks the same chat-completions wire format, so a single base-URL setting covers them all and also covers local Ollama and vLLM.

Additional context

The #1809 review called the base-URL approach "the right shape" and asked for one fix: setting only some of the variables fell through silently and could caption through Gemini, the provider the user is trying to avoid. That PR was closed for merge conflicts and inactivity, before Vertex support (#3561) landed in the same function.

To check that the request shape works outside OpenRouter, I sent exactly what the OpenRouter branch builds today (one image_url data URI, the storyboard prompt, max_tokens: 500) to iFlytek Astron's Token Plan, an OpenAI-compatible endpoint (https://maas-token-api.cn-huabei-1.xf-yun.com/v2). The image was docs/images/image-thumbnail-evidence/after.png.

None of the vision models spent the 500-token budget on reasoning. The text-only row is why the docs should say to choose a model that accepts image input.