refactor: implement _summary_list_field function to streamline handling of summary fields in coach summary; enhance payload processing and validation for string and list types

refactor: streamline evidence generation in coach summary; enhance JSON output contract and improve tokenization handling for local chat

refactor: add functions for JSON payload extraction and validation in coach summary; streamline summary key handling and improve error reporting

refactor: introduce compact representation functions for metrics, issues, and knowledge cards; enhance summary evidence generation and improve error handling in JSON extraction

refactor: enhance JSON response handling in coach summary generation; update response format instructions and improve error reporting for invalid JSON outputs

chore: remove Linux-specific dependencies from pyproject.toml and requirements.txt; update README to clarify usage of Nemotron implementation for local model inference

chore: add ninja-build and build-essential to package requirements; update README and documentation to reflect changes in model configuration and local Transformers usage

chore: update base model to NVIDIA-Nemotron-3-Nano-4B-BF16 across configurations and documentation; adjust max sequence length to 2048; upgrade GitHub Actions cache to v5

refactor: enhance _encode_bt709_output and _encode_bt709_command for improved audio stream handling and code clarity

refactor: remove sample_pose_cache module and streamline pose processing in pipeline

refactor: change default delegate to CPU in MediaPipePoseBackend and update tests accordingly

refactor: update artifact URLs and improve router device handling for better classification performance

feat: add support for local Transformers inference on Hugging Face Spaces with opt-in configuration

refactor(mediapipe): simplify GPU delegate preference logic and remove unused CUDA visibility check

refactor(mediapipe): enhance MediaPipePoseBackend to prefer tasks on GPU and improve delegate handling

docs: update default coach-summary model to build-small-hackathon/pozify-coach-summary1 and clarify fallback behavior in documentation

feat(coach): disable verifier by default

feat: add coach summary LoRA config and provider plumbing

feat: prepare filtered public fitness style corpus

feat: build Pozify-native coach summary SFT datasets

feat: expand coach retrieval with equipment and goal overlays

feat: ingest real exercise dataset into knowledge artifacts

feat: add exercise dataset to card-pack transformer

feat: load external coach knowledge packs

feat(slm): add llama.cpp coach summary provider

feat(pipeline): implement sample pose caching and enhance upload handling

feat(coach): add local transformers provider

fix(hf-spaces): install mediapipe native libraries

fix(hf-spaces): avoid pickling analysis progress callbacks

feat(pose)!: use coco17 3d router features

Implement grounded coach summary with HF SLM

feat(hf): use ZeroGPU for analysis

Set default exercise router repo

perf(pose): read dense video frames sequentially

pioritize BiLSTM model

Refactor(pipeline): improve code readability and update function calls for clarity

Merge branch 'main' into NLag/issue4

refactor(mediapipe): improve code readability and add CPU delegate option

feat(report): stream analysis progress

Enhance UnknownExercise class with debug data handling and add test for manual unknown exercise in pipeline

Avoid OpenCV H264 writer fallback errors

feat(report): add annotated issue clips

Merge remote issue 4 updates

Add issue 10 training release assets

Merge branch 'main' into NLag/issue4

fix(renderer): preserve video colors in annotations

refactor(exercises): route pipeline through exercise objects

Implement frame-level issue markers

Merge branch 'main' into NLag/issue4

refactor(exercises): introduce exercise strategies