[codex] Hash-truncate long cache path segments

[hansent]

Summary

• Add deterministic hash truncation for path segments longer than the Linux NAME_MAX limit.
• Route Roboflow metadata and model artifact cache paths through the safe join helper.
• Update direct cache path users and air-gapped cache detection to use the same bounded layout.
• Add regression coverage for long instant-model slugs.

Why

Serverless deployments were seeing OSError: [Errno 36] File name too long when instant model IDs produced very long cache directory names. The issue is per path segment, so the fix canonicalizes each segment before writing or reading cache files.

Validation

• .venv/bin/python -m black --check inference/core/utils/file_system.py inference/core/cache/model_artifacts.py inference/core/registries/roboflow.py inference/core/models/roboflow.py inference/models/transformers/transformers.py inference/models/grounding_dino/grounding_dino.py inference/models/easy_ocr/easy_ocr.py inference/core/cache/air_gapped.py tests/inference/unit_tests/core/utils/test_file_system.py tests/inference/unit_tests/core/cache/test_model_artifacts.py tests/inference/unit_tests/core/registries/test_roboflow.py
• .venv/bin/python -m isort --check-only inference/core/utils/file_system.py inference/core/cache/model_artifacts.py inference/core/registries/roboflow.py inference/core/models/roboflow.py inference/models/transformers/transformers.py inference/models/grounding_dino/grounding_dino.py inference/models/easy_ocr/easy_ocr.py inference/core/cache/air_gapped.py tests/inference/unit_tests/core/utils/test_file_system.py tests/inference/unit_tests/core/cache/test_model_artifacts.py tests/inference/unit_tests/core/registries/test_roboflow.py
• .venv/bin/pytest tests/inference/unit_tests/core/utils/test_file_system.py tests/inference/unit_tests/core/cache/test_model_artifacts.py tests/inference/unit_tests/core/registries/test_roboflow.py tests/unit/core/cache/test_air_gapped.py

[PawelPeczek-Roboflow]

double op os.fsencode(path_segment)

[PawelPeczek-Roboflow]

I agree with necessity of change, but I am not sure if that solves all issues:

• what if I just have plenty of segments which in total would exceed limit on path size?
• is it required that all segments could be altered with this function - especially with our cache purging strategy, in some cases (in theory) we could land outside desired directory, no?
• would the approach from inference_models work? we leave prefix as is and just slugify model_id to be size constraint plus add hash?