feat(generation): restore the SDXL Refiner in Generate and Canvas

[joshistoast]

Summary

The SDXL Refiner had no path through webv2 Generate or Canvas, so legacy SDXL + refiner recipes could not be reproduced.

A Refiner section (SDXL main models only) offers a refiner model picker limited to sdxl-refiner models (which remain unselectable as main models), scheduler, steps, CFG, start fraction and positive/negative aesthetic scores with the legacy defaults. When a refiner is picked, the graph adds the refiner loader, refiner conditioning (aesthetic scores, prompts wired by edge), a second denoise stage from the start fraction to the end with the base stage ending at the hand-over, and routes the refined latents into decode; seamless is applied to the refiner too. Canvas image-to-image, inpaint and outpaint use start = min(refiner start, 1 − strength) and keep masks, control, IP-adapter and regional guidance on the base stage; the refiner pass receives the same gradient denoise mask. PiD + refiner is rejected. A start fraction of 1 means off. The refiner is recorded in metadata and restored by Recall All and Remix (key, then hash, then name+base); an uninstalled or ambiguous refiner is reported in the recall notice.

Two deliberate deviations from the current legacy builder were confirmed correct by review against the backend scheduler: the base pass ends at the hand-over (current legacy regressed this for text-to-image), and the refiner's style conditioning receives the prompt by edge (legacy copied an undefined value).

Related Issues / Discussions

Follows up the legacy parity audit behind #9803.

QA Instructions

From invokeai/frontend/webv2 (Node 22.14, pnpm 10.12): pnpm check:release passes. Mock fixtures now cycle main models over sd-1 / sdxl / sdxl-refiner, so the UI can be exercised against the mock (representative profile):

1. Home → Generate from text → model picker → SDXL chip → pick an SDXL fixture: a Refiner section appears before Advanced with its controls disabled.
2. Open "Refiner model": only refiner fixtures are listed; pick one: the badge shows its name and the controls enable.
3. Switch the main model to an SD 1.x fixture: the confirm dialog lists "Refiner" and the section disappears.
4. Submitted graphs cannot be inspected on the mock (no enqueue route); graph shape is covered by compile tests. A real SDXL model plus refiner is required to see output.

Review

No material findings. Reviewers compiled 59 graph shapes from the base and head commits and found non-refiner graphs byte-identical. Follow-ups applied from review: the refiner denoise now receives the gradient mask on inpaint/outpaint; one owner for the base denoise range; recall reports an unresolved refiner; the picker's empty state no longer double-messages; start fraction 1 is treated as off; legacy input caps; docs note the Canvas strength cap.

Compatibility / Rollout

New generate settings are optional with defaults and normalize on load; no project-document migration. No API changes.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

🤖 Generated with Claude Code

---

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