fix(eval): surface pending-merge branch and raw grader response in failures

[christso]

Summary

Two diagnostics gaps found during a live dogfood of AgentV v5 against a real
downstream consumer (WiseTechGlobal/WiseTechAcademy.Evals, mid-migration to
v5).

• Results auto-push conflict warning read as data loss. When
  directPushResultsWithDetails hits a genuine results-content conflict on
  the configured results branch, it already pushes the local run to a temp
  agentv/results-sync/... branch instead of dropping it (per
  docs/adr/0007-conflict-free-results-sync-without-force-push.md). But
  maybeAutoExportRunArtifacts only printed the block reason ("...resolve it
  with a GitHub pull request"), never the sync branch name or its compare
  URL. A user seeing that warning has no way to tell the results actually
  made it to a real, pushed branch. Fixed to print the pending_merge branch
  and compare URL when present.

• LLM grader parse failures discarded the raw model response. A parse
  failure only surfaced the JSON.parse error (e.g. Unexpected identifier "This"), not what the model actually said. This makes a genuinely useful
  class of failure — a grader misrouted at a non-JSON-returning target, or a
  model ignoring the JSON-only instruction — much harder to diagnose than it
  needs to be. Fixed to include a truncated raw-response preview in the skip
  reason.

How this was found

Live dogfood run against WiseTechAcademy.Evals: evals/ace-search-subset/n1.eval.yaml,
real default target (staging search CLI) + real openai LLM-rubric graders.

• First run errored with JSON Parse error: Unexpected identifier "This" /
  "They" on both LLM-rubric graders. Root cause turned out to be on the
  consumer side (their PR feat: add retry configuration with exponential backoff for AI providers #40 migration dropped grader_target: grader from
  targets.yaml, so the grader silently self-graded against the search CLI
  instead of the LLM target) — but AgentV's error message gave no hint of
  this; only the raw-response preview added here would have shown "this
  looks like a search answer, not a JSON verdict" immediately.
• The same run's results auto-push hit a real branch divergence (other
  same-day dogfood runs had already advanced agentv/results/v1), correctly
  fell back to a sync branch, but the printed warning didn't say so.

Second (fixed) run passed at 97% with the corrected consumer-side config;
its console output demonstrates the new warning:

Warning: skipping results export: Results branch agentv/results/v1 diverged from the remote and could not be auto-merged: a genuine results content conflict remains (remote dc101b77c494, local 3a022e40a675). The remote branch is unchanged and no history was rewritten; resolve it with a GitHub pull request.
Warning: results were pushed to 'agentv/results-sync/20260706T122714Z-agentv-results-v1-614154' instead — they are not lost. Merge that branch into agentv/results/v1 with a pull request: https://github.com/WiseTechGlobal/WiseTechAcademy.Evals/compare/agentv%2Fresults%2Fv1...agentv%2Fresults-sync%2F20260706T122714Z-agentv-results-v1-614154?expand=1

Full evidence (screenshots, README with root-cause writeup): agentv-private
branch evidence/wisetechacademy-evals-agentv-v5-dogfood-2026-07-06.

Validation

• bun test packages/core/test/evaluation/graders.test.ts apps/cli/test/commands/results/remote-auto-export.test.ts — 68 pass, 0 fail (includes 4 new tests: 3 for formatPendingMergeWarnings, 1 regression for the raw-response preview).
• tsc --noEmit clean for apps/cli and packages/core.
• biome check clean on all touched files.
• Live dogfood against WiseTechAcademy.Evals (see evidence branch above) — both new warnings observed exactly as designed in a real run.

Not done here (left for human review)

• Dashboard sync-status UI has the same "reads as lost" gap for its own
  agentv/results/v1 conflict banner ("Local and remote results histories
  have diverged... resolve manually"). That's a different code path
  (apps/dashboard/src/lib/project-sync-status.ts / RunSourceToolbar.tsx,
  pull-based sync) than the one fixed here (push-based eval run auto-export),
  and a real UI change rather than a message-plumbing fix. Flagging as a
  follow-up rather than expanding this PR's scope.

Not merging — leaving open for review per this session's checkpoint policy.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	bb16a8c
Status:	✅  Deploy successful!
Preview URL:	https://10d11372.agentv.pages.dev
Branch Preview URL:	https://fix-results-pending-merge-an.agentv.pages.dev

View logs