fix(search): order CLI semantic results by rerank score, matching the web

[alishakawaguchi]

https://entire.io/gh/entireio/cli/trails/1032

Problem

entire search (CLI) and entire.io (web) return different results for the same query + repo + user — different order and different totals. Root cause: the CLI's cross-cell semantic merge orders by the wrong key.

• The CLI ordered tier 0 by BM25 and tier 1 by the per-namespace retrieval Score, and its result model had no rerankScore field at all (it silently dropped it when parsing the response).
• The shared server-side reranker (Cohere, in entire-search query-serve) is identical for both clients, but the ordering that consumes the rerank score is a separate re-implementation in each client (Go in the CLI, TS in the BFF), and the CLI's copy never got the fixes the web did.
• query-serve's own docs on the Score field: "It is NOT the field results are ordered by … sorting by it undoes reranking." The CLI was sorting by exactly that. The web BFF orders both tiers by rerank score (ENT-1425 / ENT-1431); the CLI was sitting at the pre-fix behavior.

Measured on EU prod, same user/repo:

• "dependabot" on entirehq/entire.io: 118 of 119 tier-0 positions differed between CLI and web; totals 100 vs 121.
• "antithesis reranking": different use github runners #1, 15 of 18 positions differed; CLI output carried no rerankScore.

This is the same divergence class as the "wildly different results" reports.

Fix

Bring the CLI merge to parity with the web:

• add RerankScore to the search result Meta so the CLI parses and surfaces query-serve's cross-encoder score;
• order tier 0 by rerank score desc (BM25 tie-break) when every row was reranked, else the BFF's positional mixed-capability fallback (interleave by each cell's selection rank, BM25 tie-break);
• order tier 1 by rerank score desc (retrieval Score tie-break);
• tests assert rerank ordering with BM25/Score set to the opposite order, so a regression to the old keys fails, plus a mixed-capability fallback test.

Repos, tier-2 promotion, dedupe, caps, and counts are unchanged.

Verification

• go build ./cmd/entire/..., go vet, and the search test suites pass.
• End-to-end: the patched binary's top results for "dependabot" now match the web's rerankScore order exactly (0.6276, 0.6036, 0.5790, 0.5637, …), and rerankScore is now present on every reranked row.

Follow-up (not in this PR)

• Tier-2 handling still differs (the web hard-drops tier-2; the CLI promotes/keeps it). Dormant today since query-serve no longer emits tier-2, but worth reconciling.
• The real durable fix is to stop maintaining two orderers — a single shared ordering contract, or have query-serve return final merged order so neither client re-sorts.

🤖 Generated with Claude Code

---

Note

^{Cursor Bugbot is generating a summary for commit 9175836. Configure here.}

[Copilot]

Pull request overview

This PR brings the CLI’s semantic cross-cell merge ordering into parity with the web client by parsing and using query-serve’s cross-encoder rerank score, so identical queries yield consistent ordering and totals across CLI and entire.io.

Changes:

• Add Meta.RerankScore to the search result model so the CLI parses rerank scores returned by query-serve.
• Update semantic merge ordering to sort tier 0 and tier 1 by rerank score (with appropriate tie-breakers) and implement the mixed-capability tier-0 fallback to match the BFF behavior.
• Add/adjust tests to assert rerank-first ordering and the tier-0 mixed-capability fallback behavior.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
cmd/entire/cli/search/search.go	Extends search result metadata to include rerankScore from the backend response.
cmd/entire/cli/search_v4.go	Reworks cross-cell semantic ranking to use rerank score and adds tier-0 mixed-capability fallback ordering logic.
cmd/entire/cli/search_v4_test.go	Updates tier ordering expectations and adds coverage for the tier-0 mixed-capability fallback.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[alishakawaguchi]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 87b8942. Configure here.}