perf(search): PR A — nudge O(1), BM25 single-map, tier-1 keyed sort (#1 #3 #6)

[justrach]

Three behavior-preserving perf optimizations on the per-query search/rank hot path, from the perf audit and adversarially verified (produced by a 3-attempt swarm; winner independently re-verified).

Changes

#1 — mcp.zig appendSearchSymbolNudge: the bare-identifier nudge did a full O(U+S) symbol scan (searchSymbols scans the whole symbol_index and outlines on every call) just to test whether the query names an indexed symbol. Now uses findAllSymbols — the same helper codedb_symbol uses: ensureSymbolIndex + O(1) symbol_index.get, falling back to the outline scan only when !symbol_index_complete. Output is byte-identical (the nudge only checks results.len and prints the query string).

#3 — explore.zig BM25 posting loop: dropped the doc_tf hashmap. doc_best_line already accumulates .count identically per posting, so df = doc_best_line.count() and tf = entry.value_ptr.count. Two maps → one, and the redundant per-doc .get() re-lookup is eliminated. Values bit-identical.

#6 — explore.zig tier-1 sort: the comparator hashed 4 strings + touched 2 atomics per comparison. Now precomputes (count, len) per candidate once in O(C) (mirroring the existing tier0_order pattern) and sorts a struct-key array — identical predicate (count desc, len asc) and identical input order, so ranking output is unchanged. Bonus: stops the comparator corrupting ContentCache hit/miss stats via repeated atomic reads.

Validation

• zig build test — full suite green, 0 failures.
• All three are constant-factor/algorithmic wins on the per-query path; wall-time deltas should be confirmed in ReleaseFast on a large repo (Debug numbers are inadmissible; E-core variance is real), but each is justified on complexity alone.

Part of the search-latency cluster from the Fable-5 perf audit. Independent of #659.

🤖 Generated with Claude Code

[github-actions]

Benchmark Regression Report

Thresholds: 10.00% and 50,000 ns absolute delta

NOISE means the percentage threshold was exceeded, but the absolute delta was too small to fail CI.

Tool	Base (ns)	Head (ns)	Delta	Abs Delta (ns)	Status
codedb_bundle	96399	62527	-35.14%	-33872	OK
codedb_changes	10737	11047	+2.89%	+310	OK
codedb_context	296265	312167	+5.37%	+15902	OK
codedb_deps	365	347	-4.93%	-18	OK
codedb_edit	41799	41338	-1.10%	-461	OK
codedb_find	2776	2747	-1.04%	-29	OK
codedb_hot	24565	25622	+4.30%	+1057	OK
codedb_outline	17489	16337	-6.59%	-1152	OK
codedb_read	13218	13703	+3.67%	+485	OK
codedb_search	90131	41553	-53.90%	-48578	OK
codedb_snapshot	65321	70693	+8.22%	+5372	OK
codedb_status	9450	9190	-2.75%	-260	OK
codedb_symbol	59861	51748	-13.55%	-8113	OK
codedb_tree	13852	13458	-2.84%	-394	OK
codedb_word	12021	16911	+40.68%	+4890	NOISE

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e589b3d31a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid materializing every duplicate symbol for the nudge

When the bare query is a common symbol such as init or main in a large repo, findAllSymbols duplicates every matching path/name/detail even though this branch only checks results.len. That puts O(matches) allocation/free work in front of each text codedb_search; the previous searchSymbols call was capped with max_results = 1, so materialization was bounded. Use an existence check or capped exact lookup for the nudge.

Useful? React with 👍 / 👎.