Perf: reduce initial full-index lookup and C# extraction costs

TESTING_GUIDE.md

@@ -1,5 +1,37 @@
 # Testing Guide
 
+`Extract_CSharpStaticLambdaGate_BoundsRepeatedSameLineDeclarations` checks 64
+same-line static methods with a warmed 2 MiB allocation ceiling on both runtimes,
+including C#, Razor, Blazor and CSHTML. Preserve all identities, raw start columns,
+return types and expression-body ranges. Static constructors and generic methods
+must survive; typed, tuple-return, function-pointer-return and inferred static
+lambdas must not create phantom declarations. Run the existing #4830 query corpus
+and the full symbol/reference extractor regressions alongside this fixture.
+`Extract_CSharpMethodConfirmation_RequiresOpeningParenthesis` compares every
+symbol property with optimizations disabled and requires fewer method-confirmation
+regex attempts for parameter fragments without `(`, while real multiline headers
+still reach the regex. Keep the attempt/skip counters internal to extraction tests.
+
+Confirmed C# method lookahead coverage in `SymbolExtractorCSharpRegexProbeTests`
+shares multiline methods, empty and nonempty constructor bodies, delayed body
+tokens, trailing fields and attributed accessors across C#, Razor, Blazor and
+CSHTML. Keep the warmed 512 KiB allocation ceiling, all declaration identities,
+signatures and body ranges on net8/net9. `DatabaseTests` verifies that both
+timestamp and stat reuse reject old C# contract 18 / alias contract 1 and resume
+after stamping contract 19. The full-scan upgrade regressions remain required.
+
+Initial fresh-source lookup coverage in `AuthoritativeFreshRawBulkInsertTests`
+shares canonical/display/legacy names, nested and tied ranges, out-of-range
+candidates and multi-file controls across C#, Python, JavaScript, TypeScript,
+Java, Go, Rust, C++, Kotlin and VB. Preserve the single final three-candidate
+sort and the 32-callback budget (1,000 SQLite VM instructions each) for 64
+lookups over 128 overlapping declarations, alongside savepoint and cancellation
+regressions. Run the suite on net8/net9; timing thresholds are not used here.
+Include `FreshReferenceResolutionTests` in that run: it covers direct and
+materialized source lookup, the shared 14-parameter row shape and separate SQL
+caches. The materialized outer comparison orders by the projected `start_line`;
+each probe applies the nullable-line fallback before selecting its candidate.
+
 #5339 extends the #5332 fixture to successful `files --format count --json`
 batches with the same three-snapshot/copy-byte budget for 3/12 items, checking
 counts, freshness and authority fields. Keep metadata-present/absent controls

changelog.d/unreleased/+initial-index-declaration-probes.changed.md

@@ -0,0 +1,15 @@
+---
+category: changed
+affected:
+  - src/CodeIndex/Indexer/Symbols/SymbolExtractor.CSharpScanner.cs
+  - src/CodeIndex/Indexer/Symbols/SymbolExtractor.cs
+  - tests/CodeIndex.Tests/SymbolExtractorCSharpRegexProbeTests.cs
+---
+
+## English
+
+- **Bound declaration probes during full indexing** — C#, Razor, Blazor and CSHTML avoid rechecking static-lambda arrows and prefixes across unrelated declarations, and reject impossible method prefixes before regex matching. Static constructors, generic methods and lambda filtering retain their results, with fewer regex attempts and substantially fewer temporary allocations.
+
+## 日本語
+
+- **フルインデックス時の宣言候補の検査範囲を制限** — C#・Razor・Blazor・CSHTML で別の宣言に跨る静的 lambda の arrow・prefix の再検査を避け、成立しないメソッド prefix も regex 照合の前に除外します。static constructor・generic method・lambda 除外の結果を保ちながら、regex 試行回数と一時割り当てを減らします。

changelog.d/unreleased/+initial-index-method-lookahead.changed.md

@@ -0,0 +1,14 @@
+---
+category: changed
+affected:
+  - src/CodeIndex/Indexer/Symbols/SymbolExtractor.CSharpScanner.cs
+  - src/CodeIndex/Indexer/Symbols/SymbolExtractor.Contracts.cs
+---
+
+## English
+
+- **Bound confirmed method lookahead during indexing** — C#, Razor, Blazor and CSHTML stop merging method bodies once an accessor is ruled out, reducing temporary allocations in large files while preserving ranges and recovering affected constructors. Extractor contract 19 refreshes existing rows on the next ordinary full scan.
+
+## 日本語
+
+- **インデックス作成時の確認済みメソッドの先読みを制限** — C#・Razor・Blazor・CSHTML で accessor でないと判明した時点で body の連結を止め、大きいファイルの一時割り当てを減らします。範囲を維持し、影響を受けていた constructor も回復します。次の通常フルスキャンでは抽出契約19により既存行も更新します。

changelog.d/unreleased/+initial-index-ranked-source-probes.changed.md

@@ -0,0 +1,14 @@
+---
+category: changed
+affected:
+  - src/CodeIndex/Database/DbWriter.AuthoritativeFreshReferenceSourceLookup.cs
+  - tests/CodeIndex.Tests/AuthoritativeFreshRawBulkInsertTests.cs
+---
+
+## English
+
+- **Reduce repeated source-reference sorting during initial full indexing** — temporary name indexes now select the best containing declaration before comparing at most three candidates. All languages retain the same source identities, alias and legacy-name behavior, and transactional recovery.
+
+## 日本語
+
+- **初回フルインデックスでの参照元候補の反復ソートを削減** — 一時的な名前 index で最適な包含宣言を先に選び、最後の比較を最大3候補に制限します。全言語で参照元の同一性、別名・旧形式の名前照合、トランザクションの復旧動作を維持します。

docs/initial-index-performance.md

@@ -20,13 +20,57 @@ ranked ID is consumed. Avoiding duplicate elimination removes a temporary
 B-tree per reference while preserving same-file scoping, nested range selection,
 tie-breaking, and legacy fallback. Ordinary source repair remains unchanged.
 
+Those temporary indexes also carry containment rank. Each probe stops at its
+first containing declaration, and the final comparison sorts at most three
+candidates. A non-containing early candidate is skipped before that limit.
+This applies to every language using the fresh writer and retains file-savepoint
+rollback. A regression with 128 overlapping same-name declarations and 64 lookups
+used 520 callbacks at 1,000 SQLite VM instructions per callback before this
+change; the new budget is at most 32 callbacks on both test runtimes. This measures
+lookup work, not end-to-end speed or worst-case misses across disjoint ranges.
+
 All language-independent folded symbol/reference keys use runtime-vectorized
 ASCII validation and casing. Unicode names still normalize with NFKC and apply
 the same vendored casefold deltas, but append lowered scalar values directly
 instead of allocating intermediate strings per character. Already folded ASCII
 names retain their original string. Persisted keys, their contract version and
 runtime fingerprint are unchanged; readers and every writer use the same fold.
 
+C# declaration lookahead retains a confirmed method header while checking for
+property accessors. When the first body token rules out an accessor, it stops
+there instead of joining subsequent method bodies up to a later field. This
+avoids repeated large strings in files with many multiline methods, preserves
+method ranges and signatures, and recovers affected constructor declarations.
+The same extractor handles Razor, Blazor and CSHTML. All four language keys now
+use C# extractor contract 19, so an ordinary full scan refreshes old rows even
+when source files are unchanged. Comments and accessor attributes still receive
+the existing lookahead.
+
+C# static-lambda rejection examines only arrows that can still enclose the
+candidate name and explicit-return prefixes within the current declaration
+segment. Parenthesized tuple groups remain intact. This avoids repeated scans
+and prefix allocations across earlier members on a dense line, while preserving
+static constructors, generic methods, and typed/tuple/function-pointer lambdas.
+Razor, Blazor and CSHTML share this behavior. A warmed extraction of 64 same-line
+static methods allocated 406.5 MB before these bounds and 0.93 MB after them;
+the regression fixture includes positive controls and allows 2 MiB on net8/net9.
+This allocation budget does not assert a wall-clock speed for other inputs.
+
+Confirmed method-prefix matching also requires an opening parenthesis before
+running the regex. Parameter continuation fragments ending in `) {` cannot match
+without it; the fast rejection preserves full symbol records and is covered by
+regex-attempt counts rather than a timing threshold.
+
+On a fixed 1,554-file snapshot of this repository at `0d39d0658`, a fresh Release
+.NET 8 index on macOS ARM64 with `--parallelism 2 --memory-trace` took 76.0 seconds
+before these three changes and 68.9 seconds after them. Total managed allocations
+fell from 18.46 GB to 7.83 GB. Each run used a new database and the same source
+snapshot. These are single-run observations, not a general speed guarantee.
+All pre-existing symbol identities and extraction metadata remained identical;
+seven missed constructors were recovered, removing seven declaration-site
+reference rows and updating the corresponding overload-resolution candidates.
+Both runs completed all 1,554 files without warnings or extraction errors.
+
 ## 日本語
 
 空のデータベースに対する通常の CLI フルスキャンは、初回専用の一括 writer を
@@ -46,8 +90,46 @@ DB レイアウト、抽出範囲、トランザクション境界、取消・
 を省きつつ、同一ファイルへの限定、入れ子の選択、同順位の決定、旧形式への fallback
 を維持します。通常の参照元修復処理は変更しません。
 
+一時 index には包含範囲の順位も持たせ、各照会では参照行を含む最上位の宣言だけを
+取得します。最後に並べ替える候補は最大3件です。範囲外の候補は件数制限の前に除外
+します。初回 writer を使う全言語に適用し、ファイル単位の savepoint rollback も
+維持します。同名の重複範囲128件に対する64回の照会では、変更前は SQLite VM の
+1,000命令ごとの callback が520回、変更後の回帰テスト上限は両 runtime で32回です。
+これは照会処理量の測定であり、全体の速度や互いに離れた範囲での miss の最悪値を
+保証するものではありません。
+
 全言語共通のシンボル・参照の folded key は、ランタイムのベクトル化された ASCII
 検査と小文字化を使います。Unicode 名は既存の NFKC 正規化と casefold 差分表を
 維持しつつ、文字ごとの一時文字列を作らずに小文字化した scalar を直接追加します。
 変換済みの ASCII 名は元の文字列を再利用します。保存 key、契約バージョン、実行環境
 fingerprint は変わらず、reader と各 writer は同じ変換を使います。
+
+C# 宣言の先読みでは、プロパティの accessor を調べる間、確認済みのメソッド
+header を保持します。body の先頭 token で accessor でないと分かれば停止し、
+後続メソッドの body を末尾の field まで連結し続けません。複数行メソッドが多い
+ファイルで巨大な一時文字列の反復生成を避け、メソッドの範囲・signature を保ち、
+影響を受けていた constructor 宣言も回復します。同じ抽出器を使う Razor・Blazor・
+CSHTML を含む4つの言語 key は C# 抽出契約19を使い、通常のフルスキャンで未変更の
+既存ファイルも再抽出します。コメントや accessor 属性の先読みは維持します。
+
+C# の静的 lambda の除外判定は、対象名をまだ含み得る arrow と、現在の宣言区間内の
+明示的な戻り値型 prefix に限定します。括弧で囲まれた tuple group は維持します。
+密な1行で前のメンバーまで繰り返し走査・文字列生成する処理を避け、static constructor・
+generic method・型付き／tuple／関数ポインター型 lambda の判定を保ちます。
+Razor・Blazor・CSHTML も共通です。同一行の static method 64件のウォームアップ後の
+抽出は、この制限の前で406.5 MB、後で0.93 MBを割り当てました。回帰fixtureには
+正例の対照も加え、net8/net9とも上限2 MiBを設けます。別の入力に対する所要時間の
+保証ではありません。
+
+確認済みメソッド prefix の照合でも、regex 実行の前に開き括弧を必須とします。
+`(` を含まず `) {` で終わる引数の継続断片は照合できないため、先に除外します。
+全シンボル項目の一致と regex 試行回数で検証し、時間の閾値には依存しません。
+
+このリポジトリの `0d39d0658` 時点の1,554ファイルを固定し、macOS ARM64・Release
+.NET 8・`--parallelism 2 --memory-trace` で空DBから計測したところ、3つの変更前は
+76.0秒、変更後は68.9秒でした。managed の総割り当て量は18.46 GBから7.83 GBに
+減りました。毎回新しいDBと同じソースを使った単回の観測値であり、一般的な速度を
+保証するものではありません。既存シンボルの識別情報・抽出メタデータを維持し、
+欠落していた constructor 7件を回復して、宣言位置の参照7行を除去し、対応する
+overload の解決候補も更新しました。
+両方とも全1,554ファイルを警告・抽出エラーなしで完了しています。