[cristiangirlea]

One dependency refresh for the three model-wheel bumps Dependabot proposed as manifest-only changes (#94 numpy 2.5.4, #95 onnxruntime 1.31.0, #96 tokenizers 0.23.3), with the regenerated requirements-model.lock, which those pull requests left stale. Supersedes #94, #95 and #96, which are closed once this merges.

The lock was resolved inside the pinned python:3.12-bookworm image for CPython 3.12 on Linux x86_64 and aarch64, every hash from pip hash -a sha256, platform wheels with both hashes. The fresh resolution brings huggingface-hub 2.3.0 (tokenizers allows <3.0) with the httpx2 stack in place of requests, and onnxruntime 1.31 no longer needs coloredlogs or sympy; filelock, fsspec, idna, protobuf and hf-xet moved with them. The fetcher does not use the hub library (fixed URLs, its own hash-checked layout), so nothing in it changed.

One behaviour change in the runtime itself: ONNX Runtime 1.31 starts a telemetry uploader to Microsoft's collector at import, before any code could call its off switch (a probe container with network access held an established connection to the collector within forty seconds of importing it). The service now sets ORT_DISABLE_TELEMETRY=1 in the environment every child starts with, overriding any inherited value, and a test reads it back from the child's /proc/<pid>/environ (seen failing before the fix). The security guide names the safeguard next to the rest of the child environment, and the CI step that imports the runtime straight from the image passes the same variable. The output names, inputs and the int8 parity check are unchanged.

Files: requirements-model.txt, requirements-model.lock, worker/ai_bridge/jobs.py, tests/python/test_jobs.py, docs/security.md, .github/workflows/tests.yml, CHANGELOG.md.

Checks on this head: fetcher from an empty cache (20 wheels under the new hashes, 4 models verified), model-test, model-integration and model-bench (rerank 512 documents 0.21 s, embed 32 texts 0.16 s, redact 2000 characters 0.68 s: the documented figures), python-tests (416), the full deterministic sequence of AGENTS.md, .github/scripts/audit.sh (nothing), and both worker images built for linux/arm64 under QEMU; the aarch64 wheels were hashed from PyPI but not run on an aarch64 machine. Review: independent read by Codex (gpt-6.1-sol) of the dependency graph against the wheels' metadata, the hub compatibility, the trust changes of the new HTTP stack and the telemetry switch; its two findings (the CI image import step, the missing sentence in the security guide) are fixed in the commit.