feat(results): slim index.jsonl target_execution/transcript_summary to sidecars

[christso]

Summary

Part of av-cpl5 (results-artifact-layout tracker). Implements av-cpl5.3: slim .internal/index.jsonl rows so they stay a Dashboard-ready manifest without inlining heavy sidecar payloads.

• IndexArtifactEntry.target_execution (full TargetExecutionWire envelope — command argv, complete stdout/stderr logs, transcript) is dropped from every row. target_execution_path (unchanged, already written to target-execution.json) remains the sidecar pointer. A new compact target_error_kind scalar (row-level and per-sample, in samples[]) preserves the one bit of headline signal (error_kind) that was previously read out of the inlined envelope, without requiring a sidecar read.
• IndexArtifactEntry.transcript_summary and TrialResultArtifact.transcript_summary (unbounded files_read/files_modified/shell_commands/web_fetches/errors lists) are dropped. transcript_path (unchanged) still points to the full transcript; each sample's result.json sidecar (referenced by sample_path) still carries transcript_summary in full, and apps/cli/src/commands/results/serve.ts already has a fallback (objectField(trial, 'transcript_summary') ?? objectField(runResult, 'transcript_summary')) that reads it from there when absent inline — verified this fallback already covers both old (fat) and new (slim) row shapes with no server changes needed.
• Kept projection_identity inline as-is (not moved to a sidecar): writeArtifactsFromResults reads projection_identity.id back off previously-written index.jsonl rows on disk (readExistingIndexRecords → existingRecordsByProjectionIdentity) to decide skip/update/error duplicate-policy across separate agentv eval invocations appending to the same run. Moving it to a sidecar would require an N-file read on every append to reconstruct dedup state. It's already compact (fixed set of scalar fields + optional short issues list), so it doesn't need the same slimming as target_execution/transcript_summary.
• samples[] (per-trial rollup) keeps its existing compact shape (sample_path, score, status, scores, error, cost_usd, execution_status, failure_stage, failure_reason_code) plus the new target_error_kind.

Handoff notes for av-cpl5.1 / av-cpl5.4

New/changed field names for the docs and dashboard-reader beads to build against:

• target_error_kind?: string — new, row-level and per-samples[]-entry. Replaces reading target_execution.error_kind / target_execution.errorKind directly off the row (see apps/dashboard/src/components/ResultTable.tsx's targetErrorKind() helper, which still reads the old nested shape — that helper needs updating in av-cpl5.1 to read target_error_kind instead; full target_execution is no longer present on rows/samples).
• target_execution — removed from IndexArtifactEntry/TrialResultArtifact. Full detail: target_execution_path sidecar (unchanged path/shape).
• transcript_summary — removed from IndexArtifactEntry/TrialResultArtifact. Full detail: each sample's result.json (path = <result_dir>/<sample_path>/result.json); serve.ts already falls back there when missing inline.
• projection_identity — unchanged/still inline (see rationale above).

Test plan

• bun test apps/cli/test/commands/eval/artifact-writer.test.ts (74 pass) — updated fat-row assertions to match the new slim shape.
• bun test packages/core/test/evaluation/target-execution-artifacts.test.ts (1 pass) — updated to assert target_error_kind inline and full envelope only via the target_execution_path sidecar.
• bun test apps/cli/test/eval.integration.test.ts -t "rejects removed --export" (1 pass) — updated to check transcript_summary lives in the sample's result.json, not the index row.
• bun test apps/cli/test/commands/results/serve.test.ts (114 pass, unchanged) — confirms the dashboard-serving backward-compat fallback for transcript_summary (old fat rows vs. new slim rows) already works with no server-side changes.
• bun test packages/core/test/evaluation/ (2253 pass) and apps/cli/test/commands/eval/ apps/cli/test/commands/results/ (437 pass + 1 pre-existing unrelated failure in bundle.test.ts, confirmed failing identically on an untouched sibling worktree at the same base commit — unrelated to this change).
• bun run typecheck clean for both packages/core and apps/cli.
• Live dogfood (real Azure OpenAI LLM target + real Azure LLM-rubric grader, --threshold 0.8):
  • examples/features/rubric/evals/suite.yaml test summary-task against llm target → 100% pass; inspected .internal/index.jsonl: row has no target_execution/transcript_summary, all existing sidecar path fields intact.
  • examples/features/local-cli/evals/suite.yaml test cli-provider-echo against a real CLI-provider agent target (mock CLI subprocess) graded by the same live Azure grader → 100% pass; confirmed target_execution_path sidecar (target-execution.json, 4.5KB with full command/stdout/stderr/transcript) exists and is loadable, while the index row itself stays slim (whole row ~3KB) with no inlined envelope; confirmed result.json sidecar still carries the full transcript_summary.

Closes av-cpl5.3 (leaving the bead open per process — coordinator closes after review/merge).

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	0c25bee
Status:	✅  Deploy successful!
Preview URL:	https://843f17b2.agentv.pages.dev
Branch Preview URL:	https://cpl5-index.agentv.pages.dev

View logs