test(review): add reviewer evals and the 21k expression check

[zfarrell]

Adds behavioural evals for the reviewer, plus a static check for the limit that caused today's second outage.

Evals. Eight scenarios in tests/eval/ each become a real pull request in hotdata-dev/pr-review-eval (new, private), with a copy of the reviewer workflow committed into the head branch — so the graded thing is the real file on the real pull_request path, taken from the PR under review rather than from main. That is the only way a prompt change gets pre-merge exposure, since the workflow resolves the prompt from main. Grading reads PR state through the API rather than the action's execution log, which is what keeps the reviewer swappable: reviewer_workflow and reviewer_login are the only things tying it to Claude. Reporting-only, passing on a rate out of repeats.

21k expression check. A block scalar containing an interpolation compiles to one format(...) expression capped at 21,000 characters; over it, Actions refuses to start the workflow — zero jobs, no check, org blocked. Verified against both commits: fails at 24,860 on 1f12c05, passes at 20,545 on 8b04393. Current main has 455 characters of headroom, so it warns from 90%, and separately warns about a long block with no expression yet.

Complementary to #27 rather than overlapping: #27 catches startup failure dynamically for any cause, this warns before you spend the commit. No file overlap.

Two things deferred to avoid colliding with #27, which is open on the same two workflow files: the verdict label in TOOL_USAGE_JQ, and moving the two interpolations in the context step into env: (which would remove the 21k cap from that block entirely and let #26's hardening re-land).

[github-actions]

Automated review unavailable (Claude step failed). Please review manually.

[github-actions]

Reviewer eval

No results were produced. Every scenario job failed before grading — treat this as
an eval failure, not as a passing reviewer.