dashboard(results): adapt target_error_kind reader to slim index rows - #1701

Summary

Implements av-cpl5.1 (leaf of parent tracker av-cpl5): adapt Dashboard/CLI result readers to the slim summary.json/index.jsonl shapes shipped in a983924e (cases→tests rename) and 16a21b33 (target_execution/transcript_summary → sidecars).

Investigation found most readers (manifest.ts, export.ts, projection-bundle.ts, summary.ts, report.ts) never referenced the removed cases[]/total_cases/total_instances summary fields or the row-level target_execution/transcript_summary fields directly, so no changes were needed there. summaryRunId() in core already has the run_id root/metadata fallback. The trial-level (samples[]) transcript_summary sidecar fallback in serve.ts's buildRepeatTrialReadModels already covered the new schema (reads result.json when absent inline).

The one real gap: Dashboard's ResultTable.tsx targetErrorKind() helper only read the old nested target_execution.error_kind/errorKind shape, and serve.ts's row read-model never actually passed a row-level target-error signal through to the client at all (for non-repeat rows) — so this was silently broken even before the schema change.

• apps/cli/src/commands/results/manifest.ts: type target_error_kind (new) and legacy target_execution (fat-row fallback) on ResultManifestRecord.
• apps/cli/src/commands/results/serve.ts: add recordTargetErrorKind() compat adapter and wire it into attachRunDetailReadModelFields so results[].target_error_kind is populated from either the new compact field or the legacy nested envelope.
• apps/dashboard/src/lib/types.ts: add target_error_kind to EvalCaseTrial/EvalResult.
• apps/dashboard/src/components/ResultTable.tsx: targetErrorKind() now prefers target_error_kind, falling back to the legacy nested shape.
• Tests: new/legacy coverage in serve.test.ts (API-level) and ResultTable.test.tsx (render-level).

Verification

• bun run build && bun run typecheck — clean.
• node_modules/.bin/biome check on touched files — clean.
• Focused tests: apps/cli/test/commands/results/** (249 pass), apps/dashboard/src/lib/result-table.test.ts + ResultTable.test.tsx + transcript-timeline.test.tsx (26 pass).
• Live dogfood: ran a real eval (examples/features/rubric/evals/operators.eval.yaml) against a live Azure OpenAI target with a live llm-rubric grader. Confirmed the produced summary.json has the slim shape (tests[], total_tests/passed_tests/failed_tests, no metadata.run_id) and .internal/index.jsonl rows have no target_execution/transcript_summary. Verified agentv results show/summary/report all work against the bundle.
• Green (Dashboard, new slim shape): started agentv dashboard, confirmed /api/runs and /api/runs/:id render the live bundle correctly.
• Green (Dashboard, legacy compat): hand-built a synthetic legacy fat-shape bundle (cases[] summary + row-level target_execution.error_kind) and confirmed /api/runs/:id correctly derives target_error_kind: "timeout" via the new compat fallback.
• Red (pre-fix): before the serve.ts/ResultTable.tsx change, this same legacy-bundle scenario returned no error-kind signal to the client at all — confirmed by inspecting attachRunDetailReadModelFields prior to the fix.

Test plan

• bun run build
• bun run typecheck
• bun test apps/cli/test/commands/results/ (249 pass)
• bun test apps/dashboard/src/lib/result-table.test.ts apps/dashboard/src/components/ResultTable.test.tsx apps/dashboard/src/components/transcript-timeline.test.tsx (26 pass)
• Live dogfood eval + CLI results show/summary/report
• Dashboard serve API verified against both new slim and synthetic legacy bundles

🤖 Generated with Claude Code