feat: paginate /models in UI + "Fetch All Pages" option

[dawidope]

Problem

GET /models is paginated, and the tester had no way to reach past the first page.

The registry entry for List Models had params: [], so the form offered no fields at all — every call was a bare GET /models, which returns 25 models, page 1. There was no way to ask for page 2.

Raising the limit alone does not fix this. Verified against the live API:

Query	Result
per_page=100	ignored → still per_page: 25
limit=100	clamped server-side → per_page: 50
limit=50	works — the hard ceiling
page=2	works

So 50 is a server-side cap, and reading more than 50 models always requires walking pages.

Changes

• List Models endpoint — added page and limit fields, plus a _fetchAll ("Fetch All Pages") checkbox that walks every page and merges the results. page/limit hide via visibleWhen while it is on, since the walker owns them in that mode.
• src/lib/pagination.ts (new) — extracted the page walker so it is not duplicated. /api/models already paginated with its own inline copy; it now uses the shared helper. The deAPI quirks above are documented there.
• /api/proxy — handles _fetchAll generically (any GET endpoint, never a price-calc), not as a hardcoded special case for /models. If deAPI adds another paginated GET, it only needs a registry field. The flag is stripped from params so it is never forwarded as a query param.

Note on the merged response

It carries a marker so it is not mistaken for a verbatim API response:

"_tester": { "merged_pages": true, "pages_fetched": 6, "items": 27, "truncated": false }

This matters because meta in a merged response comes from the last page (so current_page === last_page), which would otherwise look like a bug. truncated flags a stop at MAX_PAGES (20 × 50 = 1000 items) rather than silently dropping items.

Scope note

The model dropdowns were never affected — /api/models has walked all pages since 97ed2c4. This PR fixes the manual tester endpoint, and de-duplicates the walker along the way.

Testing

At 27 models everything fits in one page of 50, so the multi-page path had never actually executed. I temporarily lowered PAGE_LIMIT to 5 to force a real 6-page walk, then restored it:

Case	Result
_fetchAll=true, PAGE_LIMIT=5	27 models, 27 unique — pages_fetched=6, truncated=false ✅
_fetchAll=true, PAGE_LIMIT=50	27 models, pages_fetched=1 ✅
_fetchAll=false, page=2&limit=10	10 models, meta.current_page=2, no _tester — old path untouched ✅
invalid _endpointId + _fetchAll	400, no crash ✅
/api/models (cache)	27 models ✅

tsc --noEmit clean, npm run lint reports no warnings or errors.

🤖 Generated with Claude Code