docs: remove studio link and add paradigm info to API tab

docs: add paradigm explanations under tabs

docs: add cURL equivalent and rename Swagger tab

Fix Phase Attention Benchmark to simulate GQA 2 KV heads

Add torch.inference_mode to stop backward graph creeping

Fix Phase Attention Benchmark to simulate 24 layers

feat: Add True O(1) Phase Attention benchmark endpoint

feat: add real empirical benchmark harness with /run_benchmark endpoint

docs: update Swagger tab with complete PowerShell, Python, JS, and cURL 2-step REST examples

fix: configure clean chat template and hybrid RIF sliding window on T4 GPU

fix: add system prompt, top_p, and repetition_penalty to prevent small model loop

fix: restore native Gradio 5 launcher for ZeroGPU prestart hooks

feat: add FastAPI server with Live Swagger UI, OpenAI endpoints, and CORS

feat: upgrade holographic resolution to 2048 bands on NVIDIA A100

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

feat: integrate KalpanaDynamicCache into all 24 layers of Qwen2.5 on ZeroGPU

fix: resolve AttributeError on llm.DEFAULT_MODELS in Gradio UI startup

fix: restore list_models optional request parameter

fix: update list_models request signature

fix: define _RATE_WINDOWS dict to prevent server 500 error

fix: make _get_client_ip safe against None request

feat: GET /v1/models returns ONLY user-registered provider models

feat: enforce registered provider keys and remove HF serverless fallbacks

fix: add 429 rate limit formatting and boilerplate chunk filtering for story retrieval

fix: correct cost comparison formula based on full uncompressed document size vs compressed RIF tokens

feat: expand RIF retrieval to 1600 tokens and enable natural synthesis system prompt

fix: auto-map gemini model to gemini-2.5-flash

fix: import time inside generate method

fix: define t0 at top of generate method

fix: report registered provider key errors explicitly in chat response

fix: route gemini and registered model IDs to custom provider API caller

debug: expose last_err in RIF context fallback

fix: add browser User-Agent header to HF router requests for live LLM inference

fix: clear custom provider fallback on error and add /v1/providers/reset endpoint

feat: add Google Gemini support and global latest_provider fallback