feat([issue-3231]): pin agy imagegen to a cheap agent model and record the fixed image model

[atomantic]

Summary

Phase 1 of the render-defaults epic (Refs #3231): give Agy image renders the same cheap-tier default Codex renders already have, and stop the Settings UI from implying an image-model choice that doesn't exist.

• AGY_IMAGEGEN_DEFAULT_MODEL = 'gemini-3.5-flash-low' (server/services/imageGen/modes.js, next to the Codex pins). An unpinned agy render previously resolved to the ANTIGRAVITY_CONFIGURED_DEFAULT sentinel, which resolveCliModel maps to no --model flag at all — so agy ran whatever session model its own config selected (potentially claude-opus-4-6-thinking) just to relay one generate_image tool call. CLOUD_PROVIDER_SPECS[AGY] now falls back to the concrete pin; an explicit imageGen.agy.model in Settings still wins, as does a per-render cloudModel. Code-level default (no migration), per the Codex precedent. Agy bakes effort into the model id (-low/-medium/-high), so no separate effort pin is needed.
• AGY_IMAGEGEN_IMAGE_MODEL = 'imagen-3.0-generate-002' — the image model behind agy's generate_image tool, fixed server-side by Antigravity. All three selection channels were empirically probed and closed (no tool parameter; --model imagen-* errors pre-generation; prompt directives are ignored — agy itself falsely claims two of those routes work, which is documented at the constant so it isn't re-litigated). Render sidecars now record it as imageModel, distinct from the agent model.
• Settings → Image Gen → Agy CLI: the model field is relabeled "Agent model (drives the session — not the image model)", the placeholder names the shipped default instead of the no-longer-true "Agy's own configured default", and a read-only "Image model" row names the fixed backend. The Image Gen page's per-render picker default label follows the same resolution.

Legacy queue rows carrying the old sentinel still display "configured default" in MediaJobsQueue — correct for historical jobs. The LLM-session sentinel path (providerModels, TUI spawning) is untouched.

Test plan

• server: full suite green (23,460 passed). New/updated: resolver fallback = pin (cloudProviderConfig.test.js), sidecar records imageModel (agy.test.js), and a route-level test pinning that a fully-unpinned enqueue persists params.model === 'gemini-3.5-flash-low' into the queue job (routes/imageGen.test.js).
• client: full suite green (5,429 passed). New: ImageGenTab names imagen-3.0-generate-002 read-only and states it is not selectable.
• Manually verified end-to-end earlier: agy --model gemini-3.5-flash-low completes a generate_image render.

Refs #3231