fix: silence double toasts in Loras/media custom-catch sites (#2669)

[atomantic]

Summary

One batch of the #2669 sweep: Loras / media. Picked because it was the largest coherent remaining area that avoids the files reserved by concurrent work.

The rule applied (from the issue): a site gets { silent: true } when the call reaches the auto-toasting request() and the caller already owns its own error UI in .catch. Sites without their own error UI keep the single default helper layer.

6 sites changed across 2 files, plus a backward-compatible trailing options param threaded into 6 apiImageVideo.js wrappers (listLorasFull, getCivitaiSuggestions, deleteLoraFull, setCivitaiAuth, clearCivitaiAuth, listCachedModels).

• Loras.jsx (5) — Civitai suggestions refresh, LoRA delete, Civitai key save, Civitai key clear (all custom toast.error), plus the LoRA list load.
• MediaModels.jsx (1) — the cached-models list load.

A note on the two list-load sites

listLorasFull and listCachedModels catch into setError, which renders a persistent inline error panel (a <Banner tone="error"> / full error state) rather than a toast — so they double-report without matching a toast.error-in-.catch grep. Included them deliberately: silencing leaves the panel, which is the better of the two surfaces, and this is not a toast-less swallow (those correctly keep the single default layer and were left alone).

This isn't a new interpretation — MediaModels.handleAddFromHf already ships as addMediaModelFromHf({ url, silent: true }) paired with a setAddError catch. The two list loads were simply missed by the same sweep.

Verified-clean, no change needed

The issue's "Remaining" list called for ~9 sites here; most were already swept or are false positives. Each was checked by reading the call chain out to request():

• MusicVideo.jsx, MediaJobsQueue.jsx, MediaModels.jsx (mutations), ImageGen.jsx — already pass silent at every custom-catch site.
• removeImageWatermark / regenerateGalleryImage — hardcoded silent: true in the wrapper; the ImageGen.jsx toasts are already the only layer. Same for searchCivitaiLoras (silent = true default), which is the Loras Civitai-search toast.
• ImageClean.jsx — cleanImage / fetchCleanResult use raw fetch, never request(), so nothing auto-toasts. saveCleanResult already passes silent.
• Toast-less swallows (getCivitaiAuth().catch(() => {}), listMediaModelRegistry().catch(() => {}), deleteImage(...).catch(() => {}), listMediaJobs(...).catch(() => null)) — left alone per the rule.

Test plan

• cd client && npm test — 403 files / 4387 tests passed, 0 failures (no flakes this run).
• Grepped for tests asserting exact call args on the six touched wrappers: only useUniverseDraft.test.jsx references one (listLorasFull), as a bare vi.fn() mock with no argument assertions. All six params default to {}, so every existing caller is unaffected.

Remaining

Out-of-batch. Counts are from the issue's last scan and are not individually wrapper-verified — the stale-list problem has bitten three batches now, so re-scan before trusting them:

• Creative Director (~3) — CreativeDirector.jsx:180, PlanTab.jsx:91/103/113, OverviewTab.jsx:148
• Misc singles (~5) — CatalogCastPanel, GmailMboxImportPanel, LocalLlmTab, MessageDetail, GenomeTab, PipelineManuscriptEditor
• ManuscriptCommentCard.jsx:247,251 — generatePipelineManuscriptFix still needs an options param

Already shipped, do not re-sweep: settings + JIRA (#2670), messages/calendar/sharing, pipeline stages/canvas (#2714), writers-room/universe/universeBuilder/digital-twin/media pages/hooks (#2682), apps/GitHub/misc (#2721), and now Loras/media.

Refs #2669