@kuindji/typed-sql 0.2.0 → 0.4.0

package/src/parsing/tokenize.ts

@@ -1,103 +1,166 @@
 // Tokenization, sentinels, operators, and SQL keyword sets.
-import type { CleanIdent, CleanLooseToken, CollapseSpaces, ReplaceAll, Split, Trim, TrimPunctuation } from "./string-utils.js";
+import type { CleanIdent, CollapseSpaces, ReplaceAll, Trim, TrimPunctuation } from "./string-utils.js";
 import type { ExceedsLengthBudget, HasLineBreaks } from "./normalize.js";
 // Tokenization & parsing helpers
 
-export type Tokenize<N extends string> = CleanFilterTokens<Split<N, " ">>;
-
 // Sentinel token standing in for a TOP-LEVEL comma. It survives `MapClean`
 // (no stripped punctuation, non-empty identifier) whereas a bare `,` does not,
 // so it cleanly distinguishes a FROM-source separator from a comma nested in
 // parens / a string literal — which must still be dropped as before.
-export type CommaSep = "__tsqlcomma__";
+// A control char unrepresentable in real SQL: 1 char instead of the old
+// 13-char `__tsqlcomma__`, so every marked query string and every token list
+// it flows through interns ~14 fewer chars per top-level comma. Neutral to
+// the pipeline: not in `Punct`/`Whitespace`/`DQuotedPunct`/`OperatorToken`,
+// and `Lowercase`/`CleanIdent` leave it intact.
+export type CommaSep = "
";
 
 // Replace only TOP-LEVEL commas (paren depth 0, outside single OR double quotes)
 // with the `CommaSep` sentinel (space-padded so it tokenizes on its own). Commas
 // nested inside parens (`count(a, b)`, FROM subqueries, `insert (x, y)`, value
 // tuples), string literals, or quoted identifiers (`users as "u,1"`) are left
-// verbatim and get stripped by `MapClean` as today. The `InDString` arm tracks
-// double-quoted identifiers so a comma inside a quoted table/column alias is not
-// mistaken for a FROM-source separator. Char-walk mirrors `SplitTopLevel` /
-// `StripComments`; step-bounded.
-export type MarkTopLevelCommas<
+// verbatim and get stripped by `MapClean` as today.
+//
+// Segment-jump, not per-char (the old walk minted one growing-`Acc` string PER
+// CHARACTER on every under-budget query). Each step advances to the LEFTMOST of
+// the five state chars `,` `'` `"` `(` `)`, copying the whole run before it in a
+// single mint; inside a quote it jumps straight to the closing quote, exactly
+// like `LowercaseOutsideQuotesWorker` (`''` escapes exit+re-enter across two
+// jumps; an unterminated quote at EOF copies the rest verbatim). The `Steps` cap
+// counts JUMPS and yields `{ __c: [...] }` to the driver, so arbitrarily
+// boundary-dense inputs still complete without a partial-output bail.
+export type MarkTopLevelCommas<S extends string> =
+    string extends S
+        ? S
+        : MtcDrive<MtcWorker<S, [], false, false, "", []>>;
+
+type MtcDrive<R> =
+    R extends { __c: [infer S extends string, infer D extends any[], infer Q1 extends boolean, infer Q2 extends boolean, infer Acc extends string] }
+        ? MtcDrive<MtcWorker<S, D, Q1, Q2, Acc, []>>
+        : R;
+
+type MtcHasStruct<S extends string> =
+    S extends `${string}'${string}` ? true
+    : S extends `${string}"${string}` ? true
+    : S extends `${string}(${string}` ? true
+    : S extends `${string})${string}` ? true
+    : false;
+
+type MtcWorker<
     S extends string,
-    Depth extends any[] = [],
-    InString extends boolean = false,
-    Acc extends string = "",
-    Steps extends any[] = [],
-    InDString extends boolean = false
-> = string extends S
-    ? S
-    : Steps["length"] extends 1500
-        ? `${Acc}${S}`
-        : S extends `${infer C}${infer Rest}`
-            ? InDString extends true
-                ? MarkTopLevelCommas<Rest, Depth, InString, `${Acc}${C}`, [any, ...Steps], C extends `"` ? false : true>
-                : C extends "'"
-                    ? MarkTopLevelCommas<Rest, Depth, InString extends true ? false : true, `${Acc}${C}`, [any, ...Steps], InDString>
-                    : InString extends true
-                        ? MarkTopLevelCommas<Rest, Depth, InString, `${Acc}${C}`, [any, ...Steps], InDString>
-                        : C extends `"`
-                            ? MarkTopLevelCommas<Rest, Depth, InString, `${Acc}${C}`, [any, ...Steps], true>
-                            : C extends "("
-                                ? MarkTopLevelCommas<Rest, [any, ...Depth], InString, `${Acc}${C}`, [any, ...Steps], InDString>
-                                : C extends ")"
-                                    ? MarkTopLevelCommas<Rest, Depth extends [any, ...infer D] ? D : [], InString, `${Acc}${C}`, [any, ...Steps], InDString>
-                                    : C extends ","
-                                        ? Depth["length"] extends 0
-                                            ? MarkTopLevelCommas<Rest, Depth, InString, `${Acc} ${CommaSep} `, [any, ...Steps], InDString>
-                                            : MarkTopLevelCommas<Rest, Depth, InString, `${Acc}${C}`, [any, ...Steps], InDString>
-                                        : MarkTopLevelCommas<Rest, Depth, InString, `${Acc}${C}`, [any, ...Steps], InDString>
-            : Acc;
+    Depth extends any[],
+    InString extends boolean,
+    InDString extends boolean,
+    Acc extends string,
+    Steps extends any[]
+> = Steps["length"] extends 450
+    ? { __c: [S, Depth, InString, InDString, Acc] }
+    : InString extends true
+        ? S extends `${infer P}'${infer R}`
+            ? MtcWorker<R, Depth, false, InDString, `${Acc}${P}'`, [any, ...Steps]>
+            : `${Acc}${S}`
+        : InDString extends true
+            ? S extends `${infer P}"${infer R}`
+                ? MtcWorker<R, Depth, InString, false, `${Acc}${P}"`, [any, ...Steps]>
+                : `${Acc}${S}`
+            : S extends `${infer P},${infer R}`
+                // a structural char in the run before the first comma → it is
+                // leftmost; defer to the struct jump
+                ? MtcHasStruct<P> extends true
+                    ? MtcStructJump<S, Depth, Acc, Steps>
+                    : Depth["length"] extends 0
+                        ? MtcWorker<R, Depth, false, false, `${Acc}${P} ${CommaSep} `, [any, ...Steps]>
+                        : MtcWorker<R, Depth, false, false, `${Acc}${P},`, [any, ...Steps]>
+                : MtcHasStruct<S> extends true
+                    ? MtcStructJump<S, Depth, Acc, Steps>
+                    : `${Acc}${S}`;
 
-// Token stream for the table/alias collectors: identical to `Tokenize` except
-// top-level commas survive as `CommaSep` tokens (so `from a, b` exposes its
-// source boundary). Used ONLY by `TablesInQuery` / `AliasesInQuery`.
+// Leftmost of `'` / `"` / `(` / `)` (the caller guarantees at least one occurs
+// before any comma). Pairwise narrowing: split on a candidate; if an
+// earlier-class char appears in its prefix, that one is leftmost instead.
+type MtcStructJump<
+    S extends string,
+    Depth extends any[],
+    Acc extends string,
+    Steps extends any[]
+> = S extends `${infer P}'${infer R}`
+    ? P extends `${string}"${string}` | `${string}(${string}` | `${string})${string}`
+        ? MtcStructJump2<S, Depth, Acc, Steps>
+        : MtcWorker<R, Depth, true, false, `${Acc}${P}'`, [any, ...Steps]>
+    : MtcStructJump2<S, Depth, Acc, Steps>;
+
+type MtcStructJump2<
+    S extends string,
+    Depth extends any[],
+    Acc extends string,
+    Steps extends any[]
+> = S extends `${infer P}"${infer R}`
+    ? P extends `${string}(${string}` | `${string})${string}`
+        ? MtcStructJump3<S, Depth, Acc, Steps>
+        : MtcWorker<R, Depth, false, true, `${Acc}${P}"`, [any, ...Steps]>
+    : MtcStructJump3<S, Depth, Acc, Steps>;
+
+type MtcStructJump3<
+    S extends string,
+    Depth extends any[],
+    Acc extends string,
+    Steps extends any[]
+> = S extends `${infer P}(${infer R}`
+    ? P extends `${string})${string}`
+        ? S extends `${infer P2})${infer R2}`
+            ? MtcWorker<R2, Depth extends [any, ...infer D] ? D : [], false, false, `${Acc}${P2})`, [any, ...Steps]>
+            : `${Acc}${S}`
+        : MtcWorker<R, [any, ...Depth], false, false, `${Acc}${P}(`, [any, ...Steps]>
+    : S extends `${infer P2})${infer R2}`
+        ? MtcWorker<R2, Depth extends [any, ...infer D] ? D : [], false, false, `${Acc}${P2})`, [any, ...Steps]>
+        : `${Acc}${S}`;
+
+// String view for the table/alias collectors: identical content to plain
+// `Tokenize` input except top-level commas survive as `CommaSep` sentinels (so
+// `from a, b` exposes its source boundary). The collectors walk this string
+// DIRECTLY, word by word (the `Ct`/`Ca`/`Cn`/`Ta` scan walkers in tables.ts) —
+// the former `SplitCollectorTokens` token-ARRAY build (and the
+// collector-relevance filter that existed only to keep that array small) is
+// gone: per the round-8/9 census, every array build/destructure step minted a
+// unique-content tuple plus its apparent-`Array` types, while a word-jump
+// string walk interns its substrings and counter tuples.
 //
 // Report-scale queries (multi-line, or very long) skip the comma-marking
-// char-walk and fall back to plain `Tokenize` — the same big-query light path
+// char-walk and use the raw normalized string — the same big-query light path
 // `ValidateSQLNormalizedLightSelect` already takes. A comma cross-join in such a
 // query is negligibly rare, and avoiding the extra instantiation depth keeps the
 // largest analytics queries under the TS recursion limit.
-export type TokenizeTables<N extends string> =
+export type CollectorScanView<N extends string> =
     HasLineBreaks<N> extends true
-        ? Tokenize<N>
+        ? N
         : ExceedsLengthBudget<N> extends true
-            ? Tokenize<N>
-            : RestoreCleanFilterTokens<Split<MaybeMarkDQuotedSpaces<MarkTopLevelCommas<N>>, " ">>;
-
-export type TokenizeLoose<N extends string> =
-    RestoreCleanLooseFilterTokens<
-        Split<CollapseSpaces<RestoreWildcards<PadOperators<ProtectWildcards<MaybeMarkDQuotedSpaces<MaybeStripDQuotedPunct<N>>>>>>, " ">
-    > extends infer Toks extends string[]
-        ? N extends `${string}distinct ${string}`
-            ? DropDistinctFrom<Toks>
-            : Toks
-        : [];
-
-// `IS [NOT] DISTINCT FROM` is a comparison operator: its `from` is operator text,
-// NOT a FROM clause / table-source boundary. The column ref-scanner skips a token
-// whose `Prev` is `from` (treating it as a table source), so the RHS expression of
-// the operator (`price IS DISTINCT FROM bogus_col`) escapes validation entirely
-// (round-13 D1/D2). Drop the operator `from` — the one directly preceded by
-// `distinct` — from the token list so the RHS's `Prev` becomes `distinct`, which
-// `CanPrecedeColumn` already blesses, and the column is validated like any other.
-// `distinct` is immediately followed by the bare token `from` ONLY in this
-// operator, so the rewrite is unambiguous. The real FROM-clause `from` is untouched.
-export type DropDistinctFrom<
-    Tokens extends string[],
-    Acc extends string[] = [],
-    Prev extends string = "",
-    Steps extends any[] = []
-> = Steps["length"] extends 400
-    ? [...Acc, ...Tokens]
-    : Tokens extends [infer H extends string, ...infer R extends string[]]
-        ? H extends "from"
-            ? Prev extends "distinct"
-                ? DropDistinctFrom<R, Acc, "from", [any, ...Steps]>
-                : DropDistinctFrom<R, [...Acc, H], H, [any, ...Steps]>
-            : DropDistinctFrom<R, [...Acc, H], H, [any, ...Steps]>
-        : Acc;
+            ? N
+            : MaybeMarkDQuotedSpaces<MarkTopLevelCommas<N>>;
+
+// The collector token for one raw word of `CollectorScanView`: sentinel-restored,
+// then exactly the value the old split pushed (`TrimPunctuation<Trim<H>>`); `""`
+// means the word is punctuation/whitespace-only and never occupied a token
+// position (the old `CleanIdent<H> extends ""` empty-token filter — a non-empty
+// `CleanIdent` guarantees a non-empty `TrimPunctuation<Trim<H>>`, so `""` is a
+// safe drop sentinel). On the raw big-query path no sentinel can occur and the
+// restore is a single failed template match.
+export type CollectorToken<H extends string> =
+    ReplaceAll<H, DQuoteSpaceSentinel, " "> extends infer R extends string
+        ? CleanIdent<R> extends ""
+            ? ""
+            : TrimPunctuation<Trim<R>>
+        : never;
+
+// The padded, space-collapsed string the column ref-scanners walk DIRECTLY —
+// the string→string prefix of the old `TokenizeLoose` pipeline. The split into a
+// token ARRAY (and the separate `DropDistinctFrom` array pass) is gone: per round-8
+// census, every token-array build/destructure step minted a unique-content tuple
+// plus its apparent-`Array` types, while the word-jump string walks that replaced
+// them (`QualifiedRefScan` / `UnqualifiedRefScan` in columns.ts) intern their
+// substrings and counters. Token semantics (per-word `CleanLooseToken` transform,
+// sentinel restore, empty-token drop, `IS [NOT] DISTINCT FROM` handling) are
+// reproduced verbatim inside the scan walkers.
+export type LooseScanView<N extends string> =
+    CollapseSpaces<RestoreWildcards<PadOperators<ProtectWildcards<MaybePadModulo<MaybeMarkDQuotedSpaces<MaybeStripDQuotedPunct<N>>>>>>>;
 
 // Operator/comma characters that `PadOperators` would split on. Inside a
 // double-quoted identifier (`"u,1"`) these are part of the identifier, not
@@ -105,7 +168,7 @@ export type DropDistinctFrom<
 // ref-scan and falsely rejects an otherwise valid query. We drop them from inside
 // double-quoted spans before padding so the identifier stays a single token.
 export type DQuotedPunct =
-    "(" | ")" | "," | "=" | "<" | ">" | "+" | "-" | "*" | "/" | "|" | "&" | "!" | "?";
+    "(" | ")" | "," | "=" | "<" | ">" | "+" | "-" | "*" | "/" | "%" | "|" | "&" | "!" | "?";
 
 // Only pay for the char-walk when there is actually a double quote to handle —
 // the overwhelmingly common no-quote query short-circuits to identity.
@@ -115,24 +178,35 @@ export type MaybeStripDQuotedPunct<S extends string> =
 // Quote-aware walk that removes `DQuotedPunct` characters located INSIDE a
 // double-quoted span while leaving the quote characters and everything outside
 // the quotes untouched. `"u,1"` -> `"u1"`; `"u1".id` (no inner punctuation) is
-// unchanged. Step-bounded against runaway.
+// unchanged.
+//
+// Span-jump, not per-char: nothing outside a double-quoted span changes, so each
+// step jumps to the leftmost `"`, copies the whole preceding run in one mint,
+// finds the closing `"` and rewrites only the (short) span interior. Like the
+// old walk, single quotes are NOT tracked — every `"` toggles. An unterminated
+// `"` at EOF keeps stripping to the end (the old InDQ-at-EOF behavior).
 export type StripDQuotedPunct<
     S extends string,
-    InDQ extends boolean = false,
     Acc extends string = "",
     Steps extends any[] = []
 > = string extends S
     ? S
-    : Steps["length"] extends 1500
+    : Steps["length"] extends 300
+        ? `${Acc}${S}`
+        : S extends `${infer P}"${infer R}`
+            ? R extends `${infer Span}"${infer R2}`
+                ? StripDQuotedPunct<R2, `${Acc}${P}"${StripPunctChars<Span>}"`, [any, ...Steps]>
+                : `${Acc}${P}"${StripPunctChars<R>}`
+            : `${Acc}${S}`;
+
+// Per-char strip over a (short) double-quoted span interior only.
+type StripPunctChars<S extends string, Acc extends string = "", Steps extends any[] = []> =
+    Steps["length"] extends 200
         ? `${Acc}${S}`
         : S extends `${infer C}${infer Rest}`
-            ? C extends `"`
-                ? StripDQuotedPunct<Rest, InDQ extends true ? false : true, `${Acc}${C}`, [any, ...Steps]>
-                : InDQ extends true
-                    ? C extends DQuotedPunct
-                        ? StripDQuotedPunct<Rest, InDQ, Acc, [any, ...Steps]>
-                        : StripDQuotedPunct<Rest, InDQ, `${Acc}${C}`, [any, ...Steps]>
-                    : StripDQuotedPunct<Rest, InDQ, `${Acc}${C}`, [any, ...Steps]>
+            ? C extends DQuotedPunct
+                ? StripPunctChars<Rest, Acc, [any, ...Steps]>
+                : StripPunctChars<Rest, `${Acc}${C}`, [any, ...Steps]>
             : Acc;
 
 // Sentinel standing in for a SPACE located INSIDE a double-quoted identifier.
@@ -142,72 +216,33 @@ export type StripDQuotedPunct<
 // tokens. Marking the inner spaces keeps the identifier a single token through
 // the space-split; `RestoreDQuotedSpaces` turns each sentinel back into a real
 // space per-token before `CleanIdent`/`MapClean` runs. Mirrors `StripDQuotedPunct`.
-export type DQuoteSpaceSentinel = "__tsqldqsp__";
+// 1-char control sentinel (was the 12-char `__tsqldqsp__`) — same neutrality
+// argument as `CommaSep`.
+export type DQuoteSpaceSentinel = "";
 
 // Only pay for the char-walk when there is actually a double quote present — the
 // overwhelmingly common no-quote query short-circuits to identity.
 export type MaybeMarkDQuotedSpaces<S extends string> =
     S extends `${string}"${string}` ? MarkDQuotedSpaces<S> : S;
 
+// Span-jump sibling of `StripDQuotedPunct`: copy the run before the leftmost
+// `"` in one mint, then mark the span interior's spaces via `ReplaceAll`
+// (spans are short identifiers). Single quotes are NOT tracked — every `"`
+// toggles, exactly like the old per-char walk; an unterminated `"` keeps
+// marking to EOF.
 export type MarkDQuotedSpaces<
     S extends string,
-    InDQ extends boolean = false,
     Acc extends string = "",
     Steps extends any[] = []
 > = string extends S
     ? S
-    : Steps["length"] extends 1500
+    : Steps["length"] extends 300
         ? `${Acc}${S}`
-        : S extends `${infer C}${infer Rest}`
-            ? C extends `"`
-                ? MarkDQuotedSpaces<Rest, InDQ extends true ? false : true, `${Acc}${C}`, [any, ...Steps]>
-                : InDQ extends true
-                    ? C extends " "
-                        ? MarkDQuotedSpaces<Rest, InDQ, `${Acc}${DQuoteSpaceSentinel}`, [any, ...Steps]>
-                        : MarkDQuotedSpaces<Rest, InDQ, `${Acc}${C}`, [any, ...Steps]>
-                    : MarkDQuotedSpaces<Rest, InDQ, `${Acc}${C}`, [any, ...Steps]>
-            : Acc;
-
-// Fused token post-passes: one walk instead of the old
-// `FilterEmpty<MapClean<RestoreDQuotedSpaces<…>>>` three-walk chain. Each pass was
-// an independent element-wise map/filter, so composing them per token yields the
-// identical list (ordering preserved) while building the result spine once.
-//
-// The DQuote-space sentinel restore (`ReplaceAll<H, DQuoteSpaceSentinel, " ">`) lets
-// a quoted identifier that survived the space-split as one token (`"Order ID"`,
-// `"user alias".id`) clean to its true value. `CleanFilterTokens` is the no-restore
-// variant (plain `Tokenize`, which never marks sentinels).
-//
-// MapClean maps each token to `CleanIdent<H> extends "" ? "" : TrimPunctuation<Trim<H>>`
-// and FilterEmpty drops the `""`s. Since `CleanIdent = Lowercase<Unquote<TrimPunctuation<
-// Trim<S>>>>`, a non-empty `CleanIdent<H>` guarantees a non-empty `TrimPunctuation<Trim<H>>`,
-// so the kept value is never empty — the empty-token filter collapses to the single
-// `CleanIdent<H> extends ""` test. (The loose variant keeps an explicit empty filter
-// because `CleanLooseToken` can return `""` for a non-operator empty ident.)
-export type CleanFilterTokens<Tokens extends string[], Acc extends string[] = []> =
-    Tokens extends [infer H extends string, ...infer R extends string[]]
-        ? CleanIdent<H> extends ""
-            ? CleanFilterTokens<R, Acc>
-            : CleanFilterTokens<R, [...Acc, TrimPunctuation<Trim<H>>]>
-        : Acc;
-
-export type RestoreCleanFilterTokens<Tokens extends string[], Acc extends string[] = []> =
-    Tokens extends [infer H0 extends string, ...infer R extends string[]]
-        ? ReplaceAll<H0, DQuoteSpaceSentinel, " "> extends infer H extends string
-            ? CleanIdent<H> extends ""
-                ? RestoreCleanFilterTokens<R, Acc>
-                : RestoreCleanFilterTokens<R, [...Acc, TrimPunctuation<Trim<H>>]>
-            : never
-        : Acc;
-
-export type RestoreCleanLooseFilterTokens<Tokens extends string[], Acc extends string[] = []> =
-    Tokens extends [infer H0 extends string, ...infer R extends string[]]
-        ? CleanLooseToken<ReplaceAll<H0, DQuoteSpaceSentinel, " ">> extends infer M extends string
-            ? M extends ""
-                ? RestoreCleanLooseFilterTokens<R, Acc>
-                : RestoreCleanLooseFilterTokens<R, [...Acc, M]>
-            : never
-        : Acc;
+        : S extends `${infer P}"${infer R}`
+            ? R extends `${infer Span}"${infer R2}`
+                ? MarkDQuotedSpaces<R2, `${Acc}${P}"${ReplaceAll<Span, " ", DQuoteSpaceSentinel>}"`, [any, ...Steps]>
+                : `${Acc}${P}"${ReplaceAll<R, " ", DQuoteSpaceSentinel>}`
+            : `${Acc}${S}`;
 
 // A validation-only view of a query: blank the CONTENTS of every single-quoted
 // string literal (`'anything'` -> `''`) and mask the interior spaces of every
@@ -224,7 +259,7 @@ export type RestoreCleanLooseFilterTokens<Tokens extends string[], Acc extends s
 // Blanking the literal removes both problems at once (round-12 S1–S5). Masking
 // double-quoted spaces stops the same markers matching inside a quoted output
 // alias (round-12 A1) while leaving the identifier intact for ref validation
-// (`TokenizeLoose` restores the sentinel). The caller gates this behind a quote
+// (the ref-scan walkers restore the sentinel). The caller gates this behind a quote
 // and within-budget pre-check so report-scale queries never run the walk.
 export type ValidationScanView<S extends string> =
     S extends `${string}'${string}`
@@ -253,7 +288,7 @@ export type BlankSingleQuotedLiterals<
             : `${Acc}${S}`;
 
 export type OperatorToken =
-    | "(" | ")" | "," | "=" | "<" | ">" | "+" | "-" | "*" | "/" | "|" | "&" | "!" | "?"
+    | "(" | ")" | "," | "=" | "<" | ">" | "+" | "-" | "*" | "/" | "%" | "|" | "&" | "!" | "?"
     // `~` / `!~` are PostgreSQL regex-match operators; `[` / `]` delimit array
     // literals/subscripts. Treating them as operators makes `CanPrecedeColumn`
     // bless the RHS expression so a column ref there is validated (e.g.
@@ -264,11 +299,54 @@ export type OperatorToken =
 export type PadOperator<S extends string, Op extends string> =
     ReplaceAll<S, Op, ` ${Op} `>;
 
+// `.` control sentinel (was `.__wildcard__`) keeps the qualified `.*`
+// out of `PadOperators`' `*` padding; `` itself is never padded.
 export type ProtectWildcards<S extends string> =
-    ReplaceAll<S, ".*", ".__wildcard__">;
+    ReplaceAll<S, ".*", ".">;
 
 export type RestoreWildcards<S extends string> =
-    ReplaceAll<S, ".__wildcard__", ".*">;
+    ReplaceAll<S, ".", ".*">;
+
+// `%` is the modulo operator, but it is also the single most common character
+// inside LIKE/ILIKE pattern literals (`'%smith%'`). The plain `PadOperator`
+// chain pads EVERYWHERE — acceptable for the operators above because the
+// validation path blanks string literals first on small queries — but
+// `LooseScanView` also runs on NON-neutralized inputs (multi-line /
+// over-budget queries skip `ValidationScanView`), where padding inside a
+// literal would leak its words as blessed column candidates
+// (`'%smith%'` -> `' % smith % '` -> `smith` validated -> false reject).
+// So `%` gets its own quote-aware pad: literal interiors are copied
+// verbatim, `%` is padded only between them. `%`-free strings (the
+// overwhelming majority) short-circuit on a single pattern match.
+export type MaybePadModulo<S extends string> =
+    S extends `${string}%${string}`
+        ? S extends `${string}'${string}`
+            ? PadModuloQuoteAware<S>
+            : PadOperator<S, "%">
+        : S;
+
+// Pairwise span-jump (same shape as `BlankSingleQuotedLiterals`): hop to the
+// leftmost `'`, pad the run BEFORE it, copy the `'…'` span verbatim, recurse
+// on the tail. The `''` SQL escape pairs leftmost exactly like the blanking
+// walk. An UNTERMINATED opener pads the run before it, then copies the
+// literal tail verbatim (lenient: no padding inside what is textually a
+// string literal). Depth is the NUMBER OF
+// LITERALS, not string length; the step cap is a runaway backstop only — on
+// cap the remainder passes through UNPADDED (pre-round behavior, so a cap
+// hit can never cause a new rejection).
+type PadModuloQuoteAware<
+    S extends string,
+    Acc extends string = "",
+    Steps extends any[] = []
+> = string extends S
+    ? S
+    : Steps["length"] extends 300
+        ? `${Acc}${S}`
+        : S extends `${infer Pre}'${infer Rest}`
+            ? Rest extends `${infer Lit}'${infer After}`
+                ? PadModuloQuoteAware<After, `${Acc}${PadOperator<Pre, "%">}'${Lit}'`, [any, ...Steps]>
+                : `${Acc}${PadOperator<Pre, "%">}'${Rest}`
+            : `${Acc}${PadOperator<S, "%">}`;
 
 export type PadOperators<S extends string> =
     PadOperator<

package/src/partial.ts

@@ -16,11 +16,10 @@ import type {
     NormalizeQuery,
     SplitOnDotClean,
     SplitTopLevel,
-    TokenizeLoose,
     Trim
 } from "./parsing.js";
 import type {
-    QualifiedColumnRefs,
+    QualifiedRefScan,
     ResolveAlias,
     StripDoubleQuotes,
     TableKeysByName
@@ -76,13 +75,13 @@ export type ColumnRefValidPartialWith<
                     : true
                 : true;
 
-// Validate every qualified column ref in a token list, partial-mode.
+// Validate every qualified column ref in a fragment, partial-mode.
 export type QualifiedColumnRefsValidPartialFor<
     S extends DatabaseSchema,
     Tables extends string,
     Aliases extends string,
-    LooseTokens extends string[]
-> = QualifiedColumnRefs<LooseTokens, S, Tables, Aliases> extends infer Cols
+    RefSeg extends string
+> = QualifiedRefScan<RefSeg> extends infer Cols
     ? AllTrue<Cols extends string ? ColumnRefValidPartialWith<Cols, Tables, Aliases, S> : true>
     : true;
 
@@ -97,9 +96,7 @@ export type ValidateTableSourcePart<N extends string, S extends DatabaseSchema>
     TablesInQuery<N, S> extends infer Tables extends string
         ? AliasesInQuery<N, S> extends infer Aliases extends string
             ? AllPartTablesValid<Tables, S> extends true
-                ? TokenizeLoose<N> extends infer Toks extends string[]
-                    ? QualifiedColumnRefsValidPartialFor<S, Tables, Aliases, Toks>
-                    : true
+                ? QualifiedColumnRefsValidPartialFor<S, Tables, Aliases, N>
                 : false
             : true
         : true;
@@ -130,9 +127,7 @@ export type ValidateClausePart<Part extends string, S extends DatabaseSchema> =
     string extends Part
         ? false
         : NormalizeQuery<Part> extends infer N extends string
-            ? TokenizeLoose<N> extends infer Toks extends string[]
-                ? QualifiedColumnRefsValidPartialFor<S, never, never, Toks>
-                : true
+            ? QualifiedColumnRefsValidPartialFor<S, never, never, N>
             : false;
 
 // Scope-aware clause validation: identical to ValidateClausePart, but the
@@ -147,13 +142,12 @@ export type ValidateClausePartScoped<
     string extends Part
         ? false
         : NormalizeQuery<Part> extends infer N extends string
-            ? TokenizeLoose<N> extends infer Toks extends string[]
-                ? QualifiedColumnRefsValidPartialFor<S, Tables, Aliases, Toks>
-                : true
+            ? QualifiedColumnRefsValidPartialFor<S, Tables, Aliases, N>
             : false;
 
 // Expression-detector for a single SELECT-item token. HasSpecial covers space,
-// parens, arithmetic/comparison operators, comma, `::`, `||`. We additionally
+// parens, arithmetic operators (+ - * / %), comparison operators, comma,
+// `::`, `||`. We additionally
 // reject `[ ] " ' :` so array-indexing, quoted-with-space idents, json arrows,
 // and param/cast colons are treated as expressions (skipped, never falsely
 // rejected). A token clearing this guard is a plain identifier piece.