feat(capture): support OpenRouter as alternative vision provider for image captioning

[CodAr-man]

What this does

Adds OpenRouter as an alternative AI provider for image captioning in hyperframes capture,
alongside the existing Gemini integration.

How it works

Provider is selected at runtime based on which API key is present:

Environment Variable	Provider	Model
OPENROUTER_API_KEY	OpenRouter	google/gemma-4-26b-a4b-it
GEMINI_API_KEY / GOOGLE_API_KEY	Google Gemini	(existing behavior)
Neither	No AI captioning	(existing DOM-only behavior)

OpenRouter is prioritized if both keys are present.

Why

Not everyone has access to a Gemini API key. OpenRouter provides access to
vision-capable models (including Gemma) with a single unified API, making
this feature accessible to more users.

Backwards compatibility

✅ Purely additive — the Gemini code path is completely untouched.
✅ Users with no API keys see identical behavior to before.
✅ Verified working end-to-end on a real site capture.

[jrusso1020]

Heads up — GitHub is rendering packages/cli/src/capture/contentExtractor.ts as "Binary file not shown", which is why the diff appears empty (additions: 0, deletions: 0) even though the file is 15.8 KB.

Cause: the file was saved with a UTF-8 BOM and CRLF line endings, while the version on main is UTF-8 (no BOM) with LF. Every line then differs at the byte level even though the substantive change is small, and GitHub's diff renderer falls back to "binary". I diffed locally with -wB and the actual code change is ~30 lines and matches the description (adds the OpenRouter branch; Gemini path untouched) — so this is fixable PR hygiene, not a content problem.

To unblock review, could you:

1. Re-save the file as UTF-8 without BOM and LF line endings. In VS Code: click CRLF in the bottom status bar → LF; click UTF-8 with BOM → "Save with Encoding" → UTF-8.
2. git commit --amend && git push --force-with-lease on this branch.

Once that lands the diff will render normally and reviewers can take a real look.

One small thing for the substantive change: the rewrite dropped some of the inline comments around Gemini batching (free-tier 5 RPM vs paid-tier 2000 RPM rationale, the model-override env var note, the per-model benchmark numbers). Those are useful context for future maintainers — worth preserving in the OpenRouter version too.

Separately, I'm opening a .gitattributes PR (* text=auto eol=lf) so this class of issue can't reach a diff again — that's on the repo side, not anything you need to do.

Thanks for the contribution!

[jrusso1020]

Thanks for this @CodAr-man — the idea is good and still relevant (no alternative vision provider has landed since), so I've reimplemented it cleanly in #1478 and am closing this in favor of that.

Why a fresh PR rather than merging this one:

• The file was saved with a UTF-8 BOM + CRLF, so GitHub renders it as a binary diff (+0/-0, and the PR shows CONFLICTING) — it can't be reviewed or merged as-is.
• It only patches the image captioning loop. main has since added a second vision call site (rasterized-SVG captioning), and since the Gemini client is created only in the Gemini branch, an OpenRouter-only user would hit undefined.models and all SVG captioning would throw.
• A few smaller things: let ai: any, reusing HYPERFRAMES_GEMINI_MODEL for the OpenRouter model, no res.ok check, and a default model id (google/gemma-4-26b-a4b-it) that isn't a real OpenRouter model.

#1478 keeps your design (provider chosen by which key is present, OpenRouter prioritized, Gemini path untouched) and adds: a single captionOne dispatcher used by both the image and SVG loops, a separate HYPERFRAMES_OPENROUTER_MODEL override, a verified default (google/gemini-3.1-flash-lite), res.ok error handling, unit tests, and docs. Credited you in the PR. 🙏