deploy: Space overrides -- CUDA requirements.txt and app_space.py entry point

fix: Drop ChatInterface(type=...), removed in Gradio 6

fix: Make the Space a Gradio app, which is what ZeroGPU requires

add: Hosted generation for the Space, and retire the ZeroGPU attempt

fix: Launch the probe again, with SSR off

fix: Mount the ZeroGPU probe instead of launching a second server

add: A two-sided CUDA diagnostic for the ZeroGPU worker failure

fix: Unpin the Space's torch -- and undo a mis-targeted revert

Revert "revert: Unpin the Space's torch -- the CUDA 12 theory was wrong"

revert: Unpin the Space's torch -- the CUDA 12 theory was wrong

fix: Hold the Space's torch on CUDA 12 wheels

fix: Keep CUDA calls off the dedicated inference thread -- again

fix: Run inference inside @spaces.GPU, where the weights actually exist

fix: Don't let a half-loaded Embedder look like a loaded one

fix: Warm the retrieval embedder inside ZeroGPU's import window too

fix: Build the Space's agent at import scope, where ZeroGPU is watching

fix: Load the Space's model the one way ZeroGPU allows

fix: Accept the Space's existing CONTROLAI_BACKEND spelling, and stop guessing

fix: Unpin the Space's transformers -- the pin made the build unsolvable

fix: Serve the Space a pre-quantised checkpoint

add: A CUDA path for the demo Space, behind the existing engine contract

fix: Move the retrieval index out of git and onto the Hub

refactor: Collapse four inference backends into one MLX path

fix: Three silent retrieval bugs, and bridge in 88% of the corpus

add: Truncate conversation history to a fixed token budget

test: Try float16 instead of bfloat16 for the PyTorch CUDA path

fix: Stop the PyTorch backend from using pure greedy decoding

add: CONTROLAI_BACKEND=pytorch escape hatch to bypass the GGUF path

fix: Remove a test chunk that leaked into the shipped RAG index

fix: Ship the RAG index with the deploy -- it was never reaching the Space

fix: Use a prebuilt CPU wheel for llama-cpp-python, not a source build

fix: Wire up the GGUF deployment path and fix its tokenizer mismatch

fix: Stop final synthesis from hallucinating past a fabrication refusal

fix: Repetition-loop generation, redundant tool spam, and KaTeX placeholder leak

fix: Fall back to end-of-string match for truncated tool-call JSON

fix: Extend provenance guard to 1D arrays; repair trailing-comma JSON

fix: Don't route CUDA/ZeroGPU calls through a manually-created thread

fix: Stop silent empty-answer fallback; surface tool-error reasons

fix: Enforce parameter provenance, fix rendering/citations, add tools, harden agent loop

chore: Track scripts/upload_to_hf.py

chore: Remove dead code and stale Docker/GGUF-era artifacts

fix: Don't pass the model-holding agent as a @spaces.GPU argument

debug: Add timing/token-count diagnostics and explicit EOS to PyTorch generate()

fix: Explicitly .to('cuda') instead of device_map='auto' on ZeroGPU

perf: Use SDPA attention on CUDA instead of the eager default

fix: Raise ZeroGPU call duration budget; drop stale llama.cpp comment

fix: Launch the ZeroGPU probe Gradio app on its own port, not mounted

fix: Move the ZeroGPU probe function to module scope

fix: Register a real @spaces.GPU Gradio event handler for ZeroGPU startup check

fix: Include the actual Gradio/ZeroGPU code changes missed in the last commit