chore: align Flow dependencies and reward contract compatibility - #446

The dependency updates in #446, #448, #469 and #470 need matching locks and consumer changes. This combines those changes so they pass through one required PR gate and one main gate.

The reward worker now requires Types 0.18 and passes its complete reward contract to the scorer. Certification must satisfy the contract's digest, epsilon, delta, threshold, corpus and expiry policy. The seeded 0.95 confidence previously became a certificate delta of 0.050000000000000044 through binary subtraction, which exceeds the strict 0.05 policy. A shared decimal-complement calculation now supplies both the bound calculation and certificate. Tests still reject a delta or epsilon one floating-point value above the limit, a different threshold and a longer expiry.

The MLX extra now requires mlx-vlm 0.6.17 or later within 0.6, Transformers 5.14 through 5.16, and Jinja2 3.1 or later within 3.x. A clean installation needs Jinja for chat templates. The older allowed mlx-vlm 0.6.4 also fails in its vision tower with current MLX. The new optional Apple Silicon test runs the real Flow loader, processor and generation path with a local 39,448-parameter random model. It checks one image, two images, private scratch permissions and image cleanup after both generation and a processor refusal. It doesn't measure model quality.

The remaining changes update setup-uv to the pinned v10.0.1 commit and Ruff to 0.16.6. The workflow keeps its explicit uv 0.12.5 version and existing required gates. The README and reward guide now give the working published openadapt-flow[reward]==1.35.1 command; these source changes do not publish a new package or alter that release's artifacts or admissions.

The combined candidate passes 148 focused tests across reward certification, Types interoperability, control overlays, decisions, workflow contracts, dependency security and the VLM service. The real MLX tests pass with Transformers 5.14.0, 5.14.1, 5.15.1 and 5.16.1; the last environment also passes the 14 existing service tests. Ruff 0.16.6 passes package lint and formatting for all 588 Python files. The combined lock regenerates without a change under uv 0.12.5; release consistency also passes. The optional MLX test requires Apple Silicon and the service extra.

All original bot heads and individual reviewed implementation heads remain on preserved branches. The other three PRs will be closed as integrated only after this PR merges and its required main checks pass.