Add manual eval suite for /bills LLM features

[mikaalnaik]

Context

The /bills feature has two LLM touchpoints, both calling OpenAI gpt-5 via the Responses API, with no test or eval coverage today:

• summarizeBillText (services/billApi.ts) — summary, 8 tenet evaluations, final_judgment (yes|no|abstain), 3 Question Period questions.
• socialIssueGrader (services/social-issue-grader.ts) — binary "is this primarily a social issue" classifier.

This adds a manually-run eval suite so prompt edits, model changes, and parsing changes can be validated before shipping. It is run explicitly (pnpm eval:bills) and never wired into build, lint, or CI — LLM tokens are expensive.

What it does

• Standalone tsx runner (pnpm eval:bills) with --refresh, --only=social|analysis, --grep, and --fallback flags. No test-framework dependency added (tsx already exists for seed:bills-admin).
• Disk cache keyed by a hash of input text + prompt text (.cache/, gitignored). Editing a prompt invalidates its entries automatically; re-runs read from disk so iterating on checks/fixtures costs no tokens. --refresh forces a re-call.
• Structural checks (deterministic, gate the run): non-empty summary/steel_man, exactly 8 tenets with valid aligns|conflicts|neutral, valid final_judgment, exactly 3 non-empty QP questions with no "Mr./Madam Speaker" prefix, no Build Canada/we/our self-reference.
• Accuracy reporting (not gated — probabilistic): final_judgment vs label, social-issue confusion matrix with precision/recall.
• Cross-consistency warning: flags where summarizeBillText abstains but socialIssueGrader disagrees — surfacing the known divergence that the two use separate prompts/criteria.
• 8 hand-labeled synthetic fixtures spanning the decision space (social issue / pro-tenet economic / red-tape / administrative), plus a fetch-fixture.ts helper that bootstraps real bills via the existing fetchers + xmlToMarkdown.

Minimal edits to existing code: export SOCIAL_ISSUE_GRADER_PROMPT (for cache-keying), add the eval:bills script, gitignore .cache/.

Running

```bash
export OPENAI_API_KEY=sk-...
pnpm eval:bills # all fixtures (cached where possible)
pnpm eval:bills --refresh # bypass cache, re-call the API
pnpm eval:bills --fallback # force no-key fallback path (0 tokens)
```

Verification

• Zero-token fallback run (--fallback): harness runs end-to-end — imports resolve, fixtures load, checks execute, report prints + writes report.json, exit code gates on structural failures.
• tsc --noEmit clean; eslint clean on the new files.
• Live API runs (accuracy numbers, cache hit/refresh, prompt-regression sensitivity) are intentionally left to run manually.

🤖 Generated with Claude Code