1.0.0-beta2 - Trust hardening and a documentation site.

Sprint 6 and the post-sprint trust work, plus the first published
documentation site.

Highlights since 1.0.0-beta1:

- Judge discrimination resolution: validation now estimates the smallest
  quality difference a judge can actually see, and the optimizer refuses to
  claim a gain below that floor.
- Tool assertions can name arguments and order, and the run record carries
  each call's arguments (scalars only, truncated, credential-looking names
  redacted). The judges that read the record are shown arguments too.
- An agent-mode run says so when nothing it uses can observe behavior rather
  than text, read from a new observes_actions flag on the grader attribute.
- Per-grader score scales, so a binary judge is no longer forced onto a 0-5
  gradient.
- Documentation site at https://project.pages.drupalcode.org/ai_eval, 23 pages
  built by MkDocs, plus a first ai_eval.api.php.

Full detail in CHANGELOG.md under 1.0.0-beta2.