feat: harness engineering — tests, linting, CI, architecture enforcement

[ken-cavanagh-glean]

Summary

Adds mechanical enforcement to Seer following HES v1 principles. The eval framework that evaluates agent quality now has quality gates of its own.

• bun run check — canonical check command (typecheck + lint + test)
• Biome v2.4 for linting and formatting (2-space indent, single quotes, no semicolons)
• 77 tests across 7 files in <100ms:
  • Scoring logic (weighted average, edge cases, custom criteria)
  • CSV parsing (quoting, escaping, field handling)
  • Retry logic (5xx/429/408 retry, 4xx pass-through, network errors)
  • Criteria definitions (snapshot of all 10 defaults, lookups, scale mappings)
  • Judge prompt snapshots — locks the exact text of all 7 judge prompts
  • Architecture boundaries — enforces 5-layer import constraints mechanically
  • E2E pipeline — full judge pipeline with mocked Glean API (guidance + golden + safety + metrics + skip logic)
• GitHub Actions CI — 3-gate matrix (typecheck, lint, test) with fail-fast: false
• AGENTS.md — agent-optimized entry point (~60 lines)
• Development ledger — cross-session continuity

Key refactors (zero behavior change)

• Extracted 7 prompt builders from judge.ts → judge-prompts.ts (pure functions, snapshot-tested)
• Extracted parseCSVLine from cli.ts → lib/csv.ts
• any → unknown in API response interfaces
• Removed unused variables (DEFAULT_CONFIG in simulator, agentId in generate-agent)

Test plan

• bun run check passes (typecheck + lint + 77 tests)
• Browser walkthrough via Playwright: dashboard, golden set detail, run results, expanded rows with judge reasoning, guidance set detail, settings, new eval set
• Introduce deliberate type error → tsc --noEmit fails
• Introduce wrong-layer import → architecture test fails
• Change a judge prompt → snapshot test fails

🤖 Generated with Claude Code