fix(server): classify projector flags and align multimodal media handling #1451

Parent

• epic(server): close llama-server b10621 compatibility gaps #1431

Problem

b10621 loads a separate multimodal projector and exposes projector URL/auto/offload/device controls plus --media-path for local file:// inputs. mlxcel has direct MLX VLM checkpoints and multimodal request handling, so exact feature equivalence requires translation at the model capability and request boundary rather than accepting projector flags as no-ops.

Scope

• Classify --mmproj, --mmproj-url, --mmproj-auto, --mmproj-offload, --mmproj-device, their aliases, and environment variables.
• Translate only inert or capability-equivalent forms for integrated MLX VLM checkpoints; explicitly reject requests to load a separate GGUF projector.
• Implement compatible image/audio content-part decoding, media limits, model modality checks, and error envelopes on shared chat/responses/transcription routes.
• Align --image-min-tokens, --image-max-tokens, and --mtmd-batch-max-tokens with integrated MLX vision encoders where exact token-budget translation is possible.
• Implement --media-path only with a confined root and b10621-compatible relative file:// handling, or reject it with a supported upload alternative.
• Keep remote URL policy explicit and protect local files, loopback services, metadata services, redirects, and decompression limits.

Acceptance criteria

• Every projector and media option is classified with a parser or diagnostic test.
• Text-only models reject media before generation, while supported VLM/audio models accept compatible content-part shapes.
• Local media paths cannot traverse or escape the configured root through symlinks, encoding, or URL normalization.
• Remote media handling has request-size, decode-size, timeout, redirect, and network-address protections.
• Image-token minima, maxima, and batch limits either constrain the real encoder path or fail before serving with an architecture-specific diagnostic.
• Help does not imply that a GGUF mmproj file can be attached to an MLX checkpoint.

Validation

• Smoke-test one real VLM with an uploaded image and one audio-capable model with a short clip; include malformed and oversized media cases.