test: add LoCoMo-derived e2e acceptance cases

[knqiufan]

Summary

• Refs Tracking Issue: expand benchmark coverage #1263
• Add three pinned LoCoMo-derived acceptance cases. Reuse Harbor and powercontext-e2e acceptance; do not add a new runner.
  • locomo-temporal-banker — category 2, conv-30:q001, D1
  • locomo-multihop-football — category 3, conv-47:q026, D13
  • locomo-open-pastries — category 4, conv-44:q064, D3
• These are pinned samples, not a LoCoMo score.
• These three cases are positive recall probes only. Negative cases (abstention, conflict, or forbidden context) are out of scope here: the current scorer cannot assert forbidden_context. Those samples belong in a follow-up PR once the evaluation contract can express them. Later algorithm changes should then include both positive and negative cases on a named subset.
• The official EN/ZH sample guide is deferred until the aggregated-task path is stable, so this page does not go stale. A follow-up docs PR will use documentation-writer and humanizer, then a human pass.
• The LoCoMo README dead-link cleanup is a separate small PR: docs(locomo): remove dead CONVERSATION_PROFILE_AB link #1275.

Test plan

• make harness-check equivalent: ruff, ty, e2e/bub/tests (5 passed)
• Fill dataset.checksum from CI task_provenance_matches / replay.json harbor.task_checksum
• CI Acceptance (sqlite / oceanbase) passed after checksum backfill

AI usage statement

Used an AI coding assistant to implement the checked plan and draft this PR. Harbor checksums were taken from CI observations, not invented.

[knqiufan]

Scope update per the tracking-issue discussion: this PR now covers only the e2e sample guide and the three pinned LoCoMo-derived acceptance cases. The LoCoMo README dead-link cleanup is #1275. Ready for review.

[PsiACE]

it's lgtm, but I'm making a change to allow aggregated tasks to avoid multiple containers, so the doc could be removed from the PR for now so I can merge it.

and for docs: If AI helps draft a
page, use the documentation-writer and humanizer skills, then have a human
verify and approve the result.

[knqiufan]

Thanks @PsiACE. I dropped the EN/ZH sample guide and the CONTRIBUTING / nav / e2e README links from this PR so the three pinned cases can merge first.

I will follow up with the contributor guide after the aggregated-task path lands, using documentation-writer and humanizer plus a human pass.