Noodle post images render in the wrong art style — no style-interpretation pass (Conversation selfies have one, Noodle doesn't)

[Gunterlie]

Diagnosed by DSBwizzard (Release v2.4.2, build 28d16cb), via Prof. Mari.

Root cause

The two pipelines infer image style completely differently:

• Conversation selfies run a text-LLM pass that builds the image prompt. Its system prompt (packages/server/src/services/prompt-overrides/registry/conversation.ts:75) explicitly instructs: "Infer the appropriate art style from the character… anime/game characters should use anime/illustration style, realistic characters should use photorealistic style." The character's appearance/personality/image-instructions and the active Style Profile's guidance are fed to that LLM, so style is woven in organically. The subsequent compileImagePrompt call sets omitProfileStyleText/omitProfileSubjectTags: true to avoid duplicating it.
• Noodle post images have no equivalent LLM step. An upstream authoring LLM emits a draftPrompt; NOODLE_IMAGE_POST (packages/server/src/services/prompt-overrides/registry/noodle.ts) concatenates raw strings (draftPrompt + appearance + personality + image instructions); that goes to the image model verbatim via compileImagePrompt({ kind: "illustration" }). No art-style inference ever runs. On baseStyle: "auto" profiles the compiler contributes no style text either, so nothing pushes the image toward the character's art style.

User impact

• Anime/cartoon/stylized characters get correct-style selfies but photorealistic Noodle posts — broken visual consistency.
• Per-character characterImageInstructions are interpreted intelligently in selfies but dumped as trailing raw text in Noodle, where they compete weakly against the clean draftPrompt.
• Only workaround today is manually editing the Noodle image override to prepend art-style tags — global, no per-character granularity.

Fix (agreed direction)

Add a shared style-interpretation LLM pass at the Noodle image-service choke point (both noodle-noodler-images.service.ts and noodle-public-images.service.ts), right before compileImagePrompt — the structural mirror of the selfie pipeline. Both services already load the character context; running the pass downstream fixes every Noodle surface (public + NoodleR + generated-activity) at once, regardless of which upstream authored the draftPrompt.

• Reuses resolveIllustratorPromptRuntime, resolveImageStyleGuidanceText + formatImageStylePromptGuidance, and the selfie's omitProfileStyleText/omitProfileSubjectTags compile pattern.
• New editable prompt-override key noodle.imageInterpret carrying the art-style-inference instruction (ported from conversation.ts:75), worded to defer to an explicitly specified style so it doesn't fight a non-auto profile.
• New Noodle setting enableImageInterpretation, default on, with graceful fallback: if the pass errors/times out or the prompt connection can't resolve, compile the raw prompt as today so image generation never blocks.

Manual verification

• Anime / auto-profile character: selfie and Noodle post now match style (incl. a NoodleR post).
• Non-auto explicit profile: Noodle honors it; inference does not override it.
• Fallback: invalid prompt connection or toggle off → images still generate, one warn logged, no crash.
• Preview/review flow shows the interpreted prompt, not raw concatenation.