feat: reviewed endpoint reasoning, verifier judge keys and job filtering

[osolmaz]

This PR brings the Harbor-HF changes made during a recent benchmark campaign onto main. The first six commits are already running on the control Space, which was deployed from this branch plus #238, so merging them makes main match production. A later deploy from main without them would remove behavior that live runs depend on.

Nothing here is specific to one provider. The Pi changes let a reviewed OpenAI-compatible endpoint pass its model details (context window, reply limit, reasoning support and thinking-level map) to Harbor's Pi agent, so a run can use the highest reasoning level that endpoint accepts. The first endpoint to use this was a hosted model behind an inference binding.

The commits, in order:

• 21fba17 fix(presets): admit reviewed ACP endpoint connections
• 5c0a8c4 fix(pi): preserve reviewed endpoint reasoning metadata
• e3ffba8 fix(pi): register reviewed endpoint metadata with Harbor preflight
• 44ca670 fix(pi): preserve xhigh reasoning and reviewed token budget
• 222867e feat(inference): deliver scoped verifier judge keys to reviewed runs
• 5a0b33c fix(jobs): filter Harbor jobs before pagination
• c450f03 fix(runs): scope submission projection refresh (not deployed yet)

The judge-key change delivers a scoped key to the verifier for LLM-judged tasks. The jobs fix filters Harbor jobs before pagination, so a run's jobs are not lost among unrelated ones.

These commits were written by an agent session during the campaign and deployed without a PR, so they have not had a review. They need one before merge. The branch merges cleanly with main.

🤖 Generated with Claude Code

[osolmaz]

Replaced by #242. The Pi custom-endpoint commits (5c0a8c4, e3ffba8, 44ca670) duplicated behavior that upstream Harbor already has in 9faf4883 (#3352), so they are dropped. The remaining commits moved to #242 unchanged.