Add persistent multi-turn chat and Gemma 4 image support

[AbhinavPanwarDev]

Summary

This PR evolves the Mac app from a single-prompt interface into a persistent multi-turn chat workspace and adds local image input support for Gemma 4.

Chat experience

• Adds role-aware multi-turn conversations throughout the app and decode-service protocol.
• Persists complete chat history across chat switching and app relaunches.
• Adds a collapsible sidebar with newest chats first.
• Supports creating, selecting, renaming, and deleting conversations.
• Prevents multiple empty “New chat” pages from being created.
• Adds compact relative timestamps without continuously displaying seconds.
• Clears the composer immediately after a prompt is accepted.
• Supports Enter to send and Shift+Enter for a newline.
• Allows the latest user prompt to be edited and regenerated.
• Keeps earlier prompts immutable to avoid branching the conversation unexpectedly.
• Supports regenerating the latest response.
• Presents model load and reload actions inside every chat when generation is unavailable.

Transcript and streaming

• Renders user and assistant messages as distinct chat turns.
• Improves Markdown rendering for headings, lists, emphasis, code blocks, and inline code.
• Renders streamed output continuously instead of waiting for generation to finish.
• Reduces Markdown flickering and layout instability during streaming.
• Adds stable scroll-following and message copy actions.
• Periodically persists an active streamed response so partial output is not lost unexpectedly.
• Guards against long runs of invisible control tokens that produce no visible response.

Image support

• Adds an image picker and image preview to the Mac chat composer.
• Persists managed image attachments with their chat messages.
• Preserves attachments when switching chats, relaunching, editing, or regenerating.
• Adds image attachment metadata to app requests and decode-service IPC without embedding image bytes in history files or IPC frames.
• Adds native Gemma 4 image preprocessing and vision-tower execution.
• Adds an optional Gemma 4 vision sidecar installer and verifier.
• Streams vision weights from disk one layer at a time instead of loading the complete vision checkpoint.
• Adds 2D positional embeddings, RMS normalization, noncausal vision attention, gated GELU MLP processing, spatial pooling, standardization, and projection into the language-model embedding space.
• Adds bounded vision feature caching.
• Adds FP16 activation safeguards for the BF16-trained vision tower.
• Suppresses image/control tokens during text generation and stops pathological invisible-token loops.

Issues addressed

Closes #12 — Multi-turn chat

Closes #65 — Clear the composer after a prompt is submitted

Closes #9 — Visual Tower support

Partially addresses #61 — Support multimodal input. This PR implements image input in the Mac app and decode-service path. Audio input and multimodal requests through the OpenAI-compatible server are not included.

The TurboFieldfare core target built successfully.

Focused tests were added or extended for:

• Multi-turn request construction.
• Chat persistence and chat selection.
• Prompt editing and regeneration.
• Composer clearing.
• Decode-service protocol compatibility.
• Image attachment persistence and validation.
• Gemma 4 image preprocessing.
• Vision Metal kernels.
• Multimodal prefill behavior.
• Streaming response buffering.
• Invisible control-token handling.
• Vision sidecar planning, installation, and verification.

A complete reference-quality comparison of image output against the upstream Gemma 4 implementation has not yet been performed.

Memory and performance

The optional vision payload is stored in a separate sidecar of approximately 1.09 GiB. The complete sidecar is not loaded into memory.

Vision encoding uses bounded memory:

• Vision tensors are read directly into reusable Metal buffers.
• Only one vision layer’s weights are streamed at a time.
• Vision activation and matrix-conversion buffers are reused during an image request.
• Image preprocessing is limited by input file size, decoded pixel count, and a maximum of 280 visual tokens.
• The complete checkpoint, shard, or vision sidecar is never materialized in Swift heap memory.
• Temporary vision encoding buffers are released after the request.
• Only the projected image features required by language-model prefill remain available to the request.
• Managed image files are referenced by metadata instead of being copied into chat-history JSON or decode IPC frames.

The chat changes intentionally retain conversation text so saved chats can be restored. Active streaming uses a bounded temporary response buffer that is reset between runs.

No controlled before/after performance or peak-memory measurements were collected. Vision encoding adds temporary Metal memory and preprocessing latency only when an image is submitted.

Remaining limitations

• Audio input is not implemented.

• The OpenAI-compatible server does not yet accept image or audio content.

• Only the latest user prompt can be edited.

• Editing an image prompt regenerates the latest answer; it does not create alternate conversation branches.

• Image inference requires the optional vision sidecar to be installed.

• Image output quality still needs comparison against the upstream Gemma 4 processor and vision implementation.

• Peak memory and vision preprocessing latency still need measurement on supported Apple Silicon hardware.

• Full package-test execution requires a Swift/Xcode environment containing the necessary testing and SwiftUI preview plugins.

• The change does not load a complete checkpoint, shard, or large model tensor into Swift heap memory.

• Logs and artifacts contain no credentials, private paths, or model weights.

Validation

The following checks were used:

git diff --check

/Library/Developer/CommandLineTools/usr/bin/swiftc \
  -frontend \
  -parse \
  Sources/TurboFieldfare/Kernels/Vision/Gemma4VisionKernels.swift \
  Sources/TurboFieldfare/Runtime/Vision/Gemma4VisionEncoder.swift
  
  CLANG_MODULE_CACHE_PATH=/tmp/turbo-fieldfare-clang-cache \
swift build --disable-sandbox \
  --target TurboFieldfare \
  -Xswiftc -Xfrontend \
  -Xswiftc -interface-compiler-version \
  -Xswiftc -Xfrontend \
  -Xswiftc 6.3.2