feat(connectors): add topaz video connector (12 endpoints)

Why

Topaz Labs is live in the hosted Monid catalog but is missing from this repo. Its Video API runs the models behind Topaz Video AI as async jobs: precision and generative upscaling to 4K, denoising, motion deblur, deinterlacing, face recovery, colorization, SDR-to-HDR and frame interpolation. This PR ports the 12 video endpoints from monid-services adaptors/topaz, which is the contract.

Endpoints

All 12 post to POST /video/express and differ by the model they put in filters[]. Each one pins its own id. Rates are frames per Topaz credit, at up to 1920x1080 / above it (developer.topazlabs.com/video-models, checked 2026-10-22).

Design

• Lifecycle (provider-level, shared by all 12).
  • start posts the express body. A non-2xx answer is data with zero usage. A 2xx without a requestId throws, because there is nothing to poll.
  • poll reads GET /video/{id}/status. failed, canceled and canceling end as our 500 with the status body. On complete it reads GET /video/{id}, and returns our 502 if the record has no download link. A failing status or record read throws, because the job may still be billing.
  • stop is DELETE /video/{id}. A 2xx settles nothing; anything else is UNRESOLVED.
• Usage.
  • One pool in Topaz credits.
  • Evidence and the consolidate claim are the sum of the record's commit ledger rows. An empty ledger falls back to Topaz's own estimates.cost[0].
  • The estimate is frames ÷ frames-per-credit for the chosen model, ceiled, at least 1. The 4K rate starts where the output outgrows 1920x1080 in either orientation, so 1080x1920 is still HD.
  • The output drops the ledger and the account fields (v1's VIDEO_INTERNAL_FIELDS).
• Input.
  • v1's flat shape: video_url, the source's frame_count / frame_rate / width / height, the output size (required on upscalers), a per-endpoint model enum and knobs, and keep_audio.
  • Each endpoint's toRequest builds Topaz's express body. That body includes the encoder, profile, container and dynamicCompressionLevel; omitting the last is a 500 upstream. The express body is Topaz's internal job description, so v1's flat shape is kept.
  • Defaults are v1's, applied at the binding.
• Source URL. It must be https:// with a video extension, case-insensitive. v1's private-host rule (relayUrlIssue) rides the pattern as lookaheads, because a refinement would not survive compilation. Any numeric host that is not a canonical dotted quad is refused too: a URL parser reads 2130706433 or 0x7f.1 as 127.0.0.1.
• Timeouts. 30 min per run, as in v1. Diffusion models (Astra, Starlight) can queue upstream for most of it; the notes say so.
• Category. video-enhancement is a new leaf in connectors/categories.ts, with the hosted manifest's name and text.

Not in this PR

• The 11 Image API endpoints. They take a multipart file upload and return binary image bytes, which the engine cannot send or carry yet (#170).

Worth a look

• Fixtures are synthetic. No Topaz key was available. The seven chains use the response shapes from v1's 2026-09-16 drill (adaptors/topaz/provider.test.ts); ids, timestamps and URLs are placeholders. Recording real chains is a follow-up in tasks.md.
• One suite for the async endpoints. Like kling, the 12 share one lifecycle.test.ts driven by shared chains:
  • a literal estimate table keyed by the 12 ids;
  • the happy chain per endpoint;
  • the failure paths;
  • stop;
  • the express body;
  • schema gates with passing near-twins;
  • a live test gated on the key and TOPAZ_LIVE_SPEND=1, since it spends credits.

Verification

• deno test connectors/topaz: 15 passed, 0 failed, 1 ignored (the live test).
• Full deno test --no-check: 1313 passed, 0 failed, 222 ignored.
• deno task check: no errors in topaz. The 24 remaining errors are in the dataforseo, magic-hour and openai test files, and they reproduce on a clean main.
• deno task version:check: no contract-surface changes.
• deno task compiler:compile twice: byte-identical.
• ids.lock.json: adds exactly the 12 topaz ids.
• deno lint and deno fmt: clean.
• OpenSpec: openspec/changes/add-connector-topaz/.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added 12 Topaz video-enhancement options, including upscaling, animation upscaling, creative and generative upscaling, denoising, deblurring, artifact repair, deinterlacing, face recovery, colorization, frame interpolation, and SDR-to-HDR conversion.
  • Added support for selecting processing models and output settings, preserving audio by default, and tracking usage in Topaz credits.
  • Added a video-enhancement category for these tools.