turbine-orm 0.65.0 → 0.66.0

package/dist/cjs/cli/prisma-schema.js

@@ -264,8 +264,13 @@ function parseAttributes(fragment, line) {
         else {
             j = k;
         }
-        // `@db.VarChar(255)` etc. - keep only the head so `db` is the recorded name.
-        attrs.push({ name: rawName.split('.')[0], args, block, line });
+        // `@db.VarChar(255)` etc. The HEAD is the recorded name (so consumers keep
+        // matching on `db`), and everything after the first dot is kept as the
+        // native type rather than discarded.
+        const dot = rawName.indexOf('.');
+        const head = dot === -1 ? rawName : rawName.slice(0, dot);
+        const nativeType = dot === -1 ? undefined : rawName.slice(dot + 1);
+        attrs.push({ name: head, ...(nativeType ? { nativeType } : {}), args, block, line });
         i = j;
     }
     return attrs;
@@ -298,16 +303,28 @@ function matchBrace(s, open) {
     }
     return -1;
 }
-/** Scan the top level for `keyword Name { ... }` blocks via brace matching. */
-function scanBlocks(src) {
+/**
+ * Scan the top level for `keyword Name { ... }` blocks via brace matching.
+ *
+ * A header whose keyword is not one we handle is skipped, and now RECORDED:
+ * the switch in `parsePrismaSchema` has a `default` branch that pushes an
+ * "unsupported block" warning, but nothing could ever reach it because the
+ * filter here dropped those blocks first. The body is deliberately still
+ * re-scanned (unchanged behaviour) so a recognized block nested under an
+ * unrecognized one is not lost.
+ */
+function scanBlocks(src, warnings) {
     const blocks = [];
     const headerRe = /(^|\n)[ \t]*([a-zA-Z]+)[ \t]+([A-Za-z_]\w*)[ \t]*\{/g;
     let m;
     // biome-ignore lint/suspicious/noAssignInExpressions: standard regex exec loop
     while ((m = headerRe.exec(src)) !== null) {
         const keyword = m[2];
-        if (!BLOCK_KEYWORDS.has(keyword))
+        if (!BLOCK_KEYWORDS.has(keyword)) {
+            warnings.push(`Skipped unsupported block "${keyword} ${m[3]}" (line ${lineAt(src, m.index + m[1].length)}): the parser ` +
+                `understands model / view / type / enum / datasource / generator blocks only.`);
             continue;
+        }
         const braceOpen = src.indexOf('{', m.index);
         const close = matchBrace(src, braceOpen);
         const headerLine = lineAt(src, m.index + m[1].length);
@@ -366,23 +383,57 @@ function parseCompoundKey(attr, line) {
 function truncate(s, n = 60) {
     return s.length > n ? `${s.slice(0, n)}...` : s;
 }
+/**
+ * Field attributes that are RECOGNIZED but change what the model means, and
+ * that the name mapper does not act on. Each gets a warning naming what is
+ * being lost, so it surfaces in the report instead of being dropped in silence.
+ */
+const NOTED_FIELD_ATTRS = {
+    ignore: '@ignore excludes the field from the Prisma client. Turbine generates from the DATABASE, so the column IS present on the generated client.',
+};
+/** Same, for block (`@@`) attributes. */
+const NOTED_BLOCK_ATTRS = {
+    ignore: '@@ignore excludes the model from the Prisma client. Turbine generates from the DATABASE, so the table IS present on the generated client.',
+    schema: '@@schema (multi-schema datasource) is not represented. Turbine introspects ONE Postgres namespace per run, and two models in different schemas can resolve to the same bare table name.',
+    fulltext: '@@fulltext indexes are not represented in the generated metadata.',
+};
 /** Parse a single field declaration line. Returns null for a non-field line. */
 function parseFieldLine(text, line, warnings) {
-    // First token = field name, second token = type. Both are simple words; the
-    // type may carry a trailing `[]` and/or `?`.
-    const m = text.match(/^([A-Za-z_]\w*)\s+([A-Za-z_]\w*)(\[\])?(\?)?/);
+    // First token = field name, second token = type. The type is a simple word or
+    // an `Unsupported("...")` call, and may carry a trailing `[]` and/or `?`.
+    //
+    // `Unsupported(...)` HAS to be matched here rather than left to the bare-word
+    // branch: `data Unsupported("tsvector")?` matched only up to `Unsupported`,
+    // so the `?` sat behind an unconsumed `(` and the field recorded
+    // `optional: false`. A nullable column read as required is a required/optional
+    // INVERSION, which is exactly the kind of thing a migration report exists to
+    // catch.
+    const m = text.match(/^([A-Za-z_]\w*)\s+(Unsupported\(\s*"(?:[^"\\]|\\.)*"\s*\)|[A-Za-z_]\w*)(\[\])?(\?)?/);
     if (!m) {
         // Not a field (e.g. a stray token); skip leniently.
         warnings.push(`Skipped unrecognized line ${line}: "${truncate(text)}"`);
         return null;
     }
     const name = m[1];
-    const type = m[2];
+    const rawType = m[2];
     const isList = m[3] === '[]';
     const optional = m[4] === '?';
     const rest = text.slice(m[0].length);
     const attrs = parseAttributes(rest, line);
-    return { name, type, optional, isList, attrs, line };
+    let type = rawType;
+    let unsupported;
+    if (rawType.startsWith('Unsupported')) {
+        type = 'Unsupported';
+        unsupported = unquote(rawType.slice(rawType.indexOf('(') + 1, rawType.lastIndexOf(')')).trim());
+        warnings.push(`Line ${line}: field "${name}" is Unsupported("${unsupported}"). Prisma cannot read or write it; ` +
+            `Turbine generates the column from the database, so it IS present on the generated client.`);
+    }
+    for (const attr of attrs) {
+        const note = NOTED_FIELD_ATTRS[attr.name];
+        if (note)
+            warnings.push(`Line ${line}: field "${name}": ${note}`);
+    }
+    return { name, type, ...(unsupported !== undefined ? { unsupported } : {}), optional, isList, attrs, line };
 }
 function parseModelBody(block, kind, src, warnings) {
     const model = {
@@ -408,7 +459,14 @@ function parseModelBody(block, kind, src, warnings) {
                 else if (attr.name === 'id' || attr.name === 'unique') {
                     model.compoundKeys.push(parseCompoundKey(attr, line));
                 }
-                // @@index, @@schema, and anything else: recorded in blockAttrs, unused.
+                else {
+                    // @@index and anything else: recorded in blockAttrs, unused. The
+                    // few that change what the model MEANS get a warning rather than
+                    // silence (see NOTED_BLOCK_ATTRS).
+                    const note = NOTED_BLOCK_ATTRS[attr.name];
+                    if (note)
+                        warnings.push(`Line ${line}: model "${block.name}": ${note}`);
+                }
             }
             continue;
         }
@@ -425,7 +483,10 @@ function parseEnumBody(block, src) {
         if (text.startsWith('@@')) {
             for (const attr of parseAttributes(text, line)) {
                 if (attr.name === 'map') {
-                    const arg = attr.args.find((a) => a.key === undefined);
+                    // `@@map("x")` and `@@map(name: "x")` are both valid; reading only the
+                    // positional form silently dropped the mapped enum-type name, matching
+                    // the field-level `@map` bug in prisma-resolve.ts.
+                    const arg = attr.args.find((a) => a.key === undefined || a.key === 'name');
                     if (arg?.kind === 'string' && arg.value)
                         en.map = arg.value;
                 }
@@ -529,13 +590,19 @@ function resolvePrismaDatasourceUrl(ast, env) {
 function parsePrismaSchema(source) {
     const src = stripComments(source);
     const ast = { models: [], enums: [], datasources: [], warnings: [] };
-    for (const block of scanBlocks(src)) {
+    for (const block of scanBlocks(src, ast.warnings)) {
         switch (block.keyword) {
             case 'model':
                 ast.models.push(parseModelBody(block, 'model', src, ast.warnings));
                 break;
             case 'view':
                 ast.models.push(parseModelBody(block, 'view', src, ast.warnings));
+                // A Prisma `view` is READ-ONLY in the Prisma client. Turbine resolves it
+                // against an introspected view and emits an ordinary delegate, so
+                // create/update/delete become reachable where Prisma refused them. Said
+                // out loud, because `type` blocks already warn and this one did not.
+                ast.warnings.push(`Block "view ${block.name}" is read-only in Prisma; the generated Turbine delegate is not. ` +
+                    `Writes through it will reach the database if the view is updatable.`);
                 break;
             case 'type':
                 // Composite/embedded types (MongoDB) are not tables. Parse leniently so
@@ -550,11 +617,10 @@ function parsePrismaSchema(source) {
                 // Not a table, but it declares the connection string the CLI can reuse.
                 ast.datasources.push(parseDatasourceBody(block, src));
                 break;
-            case 'generator':
-                // Configuration block - irrelevant to name mapping.
-                break;
             default:
-                ast.warnings.push(`Skipped unsupported block "${block.keyword} ${block.name}".`);
+                // 'generator': configuration, irrelevant to name mapping. Any other
+                // keyword never reaches here (scanBlocks filters and warns), so this
+                // stays a silent catch-all rather than a second warning site.
                 break;
         }
     }

package/dist/cjs/cli/sql-statements.d.ts

@@ -0,0 +1,125 @@
+/**
+ * The ONE SQL statement tokenizer the migration tooling speaks.
+ *
+ * ## Why this module exists
+ *
+ * `migrate.ts` executes migration SQL and `destructive.ts` decides whether that
+ * SQL is allowed to run. Both need to answer the same two questions, "where does
+ * one statement end", and "which characters are code rather than comment or
+ * literal", and until v0.66 each answered them with its own hand-written lexer.
+ * They disagreed, and the one that was wrong was the GUARD:
+ *
+ *   - `destructive.ts` ended a block comment at the first `*\/`
+ *     (`sql.indexOf('*\/', i + 2)`). Postgres NESTS block comments, so a
+ *     commented-out block that itself contains a comment, the ordinary shape of
+ *     "I disabled this for now", reopened as CODE partway through and the
+ *     scanner resynchronised in the middle of the file.
+ *   - `destructive.ts` then split statements with `text.split(';')`, so a
+ *     semicolon inside a quoted identifier (`DROP TABLE "we;ird"`) cut a
+ *     statement in half and neither half matched a rule.
+ *
+ * Both bugs FAIL OPEN. The worst measured case printed a partial inventory: for
+ * a file holding a nested commented-out block, a `DROP TABLE`, and a `DELETE`,
+ * the guard listed only the DELETE. The operator confirms the inventory they
+ * were shown, and the unlisted DROP TABLE runs under that confirmation.
+ *
+ * The tokenizer that replaced them arrived with the same shape of hole one
+ * layer down, and it is worth naming the shape rather than the instances: EVERY
+ * disagreement with the server's own lexer fails open, because the guard reads
+ * a file the server will not execute and the server executes a file the guard
+ * never read. Two more were found and fixed after the rewrite, both verified
+ * dropping a table live on PostgreSQL 16: a `$` INSIDE an identifier read as a
+ * dollar-quote opener (see {@link IDENT_CONT}), and the inner semicolons of a
+ * PG14+ `BEGIN ATOMIC` routine body read as statement terminators (see
+ * {@link opensAtomicBody}). The rule the module is written to now is to copy
+ * Postgres's lexical rules exactly, and where a case is genuinely ambiguous, to
+ * pick the reading that shows the operator MORE, never less.
+ *
+ * A second lexer that agrees today drifts tomorrow, so there is exactly one
+ * here and both callers consume it. {@link tokenizeSql} does the whole job in a
+ * single pass and hands back, per statement, BOTH the verbatim source (what
+ * `migrate.ts` executes, byte for byte) and the comment/literal-stripped source
+ * (what `destructive.ts` matches its rules against). Neither view can describe a
+ * different set of statements than the other, because there is only one walk.
+ *
+ * Pure leaf: no imports at all, and in particular none from `migrate.ts` or
+ * `destructive.ts` (`migrate.ts` already imports `destructive.ts`, so the shared
+ * code cannot live in either of them).
+ */
+/** One top-level statement, in both of the forms its two consumers need. */
+export interface SqlStatement {
+    /**
+     * The statement exactly as written, trimmed, with its terminating semicolon
+     * removed. This is what gets EXECUTED, so it is sliced straight out of the
+     * source rather than reassembled: a tokenizer that rebuilds the text can
+     * silently alter it, and a mangled statement run against production is the
+     * failure this module exists to prevent.
+     */
+    raw: string;
+    /**
+     * The same statement with comments removed, string literals emptied to `''`,
+     * and dollar-quoted bodies emptied to `''`. Quoted identifiers are kept
+     * VERBATIM, because the destructive rules match on object names.
+     */
+    stripped: string;
+    /**
+     * The same statement with comments removed and EVERYTHING ELSE verbatim:
+     * string literals, quoted identifiers, and dollar-quoted bodies all keep
+     * their contents.
+     *
+     * This is the view dynamic SQL has to be read through. A `DO` block keeps its
+     * payload in a LITERAL (`EXECUTE 'DROP TABLE users'`), which is precisely
+     * what `stripped` empties, so the destructive scanner cannot use `stripped`
+     * for a procedural body and cannot use `raw` either (a comment in the body
+     * would hide the statement after it). Removing the comments here means the
+     * scanner never has to strip them itself, which is what it used to do, with a
+     * pair of regexes that were the exact hand-written lexer this module exists
+     * to delete.
+     */
+    code: string;
+    /**
+     * Executable SQL carried INSIDE this statement, in source order, with its
+     * delimiters excluded: every dollar-quoted body, plus a PG14+ `BEGIN ATOMIC`
+     * routine body. A `DO` block or a routine source is code rather than data, so
+     * the destructive scanner re-scans these.
+     */
+    blocks: string[];
+    /** True when the statement holds nothing but comments and whitespace. */
+    commentOnly: boolean;
+}
+/**
+ * Split a SQL script into its top-level statements.
+ *
+ * A semicolon inside a single-quoted string (including a backslash-escaping
+ * `E'...'` string), a double-quoted identifier, a dollar-quoted body, a PG14+
+ * `BEGIN ATOMIC` routine body, a line comment, or a block comment (which
+ * Postgres allows to NEST) does not terminate a statement. Nor does a `$` that
+ * merely continues an identifier open a dollar-quoted body, which is the same
+ * question asked from the other side (see IDENT_CONT and opensAtomicBody for
+ * what each of those two cost when they were wrong).
+ *
+ * Every returned statement is trimmed and carries no
+ * trailing semicolon; nothing is dropped, so a comment-only fragment comes back
+ * flagged rather than missing (the executor skips those, the scanner ignores
+ * them, and neither has to re-derive the fact).
+ */
+export declare function tokenizeSql(sql: string): SqlStatement[];
+/**
+ * Split a SQL script into individual executable statements on top-level
+ * semicolons, dropping comment-only fragments.
+ *
+ * This is the list `-- turbine:no-transaction` migrations issue one statement
+ * per `client.query()` call, which is the one production-destroying failure
+ * mode of the migration runner (a partial statement executed against
+ * production), so the behavior is pinned by exhaustive unit tests.
+ */
+export declare function splitSqlStatements(sql: string): string[];
+/**
+ * The whole script with comments removed, string and dollar-quoted literals
+ * emptied, and quoted identifiers preserved: the view the destructive rules are
+ * written against, joined back into one string.
+ *
+ * Statement-level consumers should prefer {@link tokenizeSql}, whose per-
+ * statement `stripped` is the same text without having to re-split it.
+ */
+export declare function stripCommentsAndStrings(sql: string): string;

package/dist/cjs/cli/sql-statements.js

@@ -0,0 +1,378 @@
+"use strict";
+/**
+ * The ONE SQL statement tokenizer the migration tooling speaks.
+ *
+ * ## Why this module exists
+ *
+ * `migrate.ts` executes migration SQL and `destructive.ts` decides whether that
+ * SQL is allowed to run. Both need to answer the same two questions, "where does
+ * one statement end", and "which characters are code rather than comment or
+ * literal", and until v0.66 each answered them with its own hand-written lexer.
+ * They disagreed, and the one that was wrong was the GUARD:
+ *
+ *   - `destructive.ts` ended a block comment at the first `*\/`
+ *     (`sql.indexOf('*\/', i + 2)`). Postgres NESTS block comments, so a
+ *     commented-out block that itself contains a comment, the ordinary shape of
+ *     "I disabled this for now", reopened as CODE partway through and the
+ *     scanner resynchronised in the middle of the file.
+ *   - `destructive.ts` then split statements with `text.split(';')`, so a
+ *     semicolon inside a quoted identifier (`DROP TABLE "we;ird"`) cut a
+ *     statement in half and neither half matched a rule.
+ *
+ * Both bugs FAIL OPEN. The worst measured case printed a partial inventory: for
+ * a file holding a nested commented-out block, a `DROP TABLE`, and a `DELETE`,
+ * the guard listed only the DELETE. The operator confirms the inventory they
+ * were shown, and the unlisted DROP TABLE runs under that confirmation.
+ *
+ * The tokenizer that replaced them arrived with the same shape of hole one
+ * layer down, and it is worth naming the shape rather than the instances: EVERY
+ * disagreement with the server's own lexer fails open, because the guard reads
+ * a file the server will not execute and the server executes a file the guard
+ * never read. Two more were found and fixed after the rewrite, both verified
+ * dropping a table live on PostgreSQL 16: a `$` INSIDE an identifier read as a
+ * dollar-quote opener (see {@link IDENT_CONT}), and the inner semicolons of a
+ * PG14+ `BEGIN ATOMIC` routine body read as statement terminators (see
+ * {@link opensAtomicBody}). The rule the module is written to now is to copy
+ * Postgres's lexical rules exactly, and where a case is genuinely ambiguous, to
+ * pick the reading that shows the operator MORE, never less.
+ *
+ * A second lexer that agrees today drifts tomorrow, so there is exactly one
+ * here and both callers consume it. {@link tokenizeSql} does the whole job in a
+ * single pass and hands back, per statement, BOTH the verbatim source (what
+ * `migrate.ts` executes, byte for byte) and the comment/literal-stripped source
+ * (what `destructive.ts` matches its rules against). Neither view can describe a
+ * different set of statements than the other, because there is only one walk.
+ *
+ * Pure leaf: no imports at all, and in particular none from `migrate.ts` or
+ * `destructive.ts` (`migrate.ts` already imports `destructive.ts`, so the shared
+ * code cannot live in either of them).
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.tokenizeSql = tokenizeSql;
+exports.splitSqlStatements = splitSqlStatements;
+exports.stripCommentsAndStrings = stripCommentsAndStrings;
+/**
+ * A dollar-quote tag. Postgres allows digits after the first character
+ * (`$do1$`), so a tag regex that stops at letters reads the body as code and
+ * misses everything inside it. The leading character can never be a digit,
+ * which is what keeps a `$1` bind placeholder from opening a quoted body.
+ *
+ * This pattern answers "is this a well-formed tag", NOT "does a body open
+ * here". The second question is decided before the pattern is ever run, by
+ * whether the walk is standing at the start of a token at all: see
+ * {@link IDENT_CONT}.
+ */
+const DOLLAR_TAG = /^\$([A-Za-z_][A-Za-z_0-9]*)?\$/;
+/**
+ * Postgres identifier characters, copied from its own lexer rather than from
+ * intuition (`scan.l`: `ident_start [A-Za-z\200-\377_]`, `ident_cont
+ * [A-Za-z\200-\377_0-9\$]`). Every non-ASCII character is an identifier
+ * character, and so is `$` in every position but the first.
+ *
+ * That `$` is the whole reason the walk consumes an identifier as ONE token.
+ * Postgres lexes by LONGEST MATCH, so in `SELECT x$y$ FROM t` the identifier
+ * rule claims all four characters of `x$y$` and the dollar-quoting rule never
+ * gets a look at that `$`. Reading `$y$` as an opener instead starts a body
+ * whose tag never appears again, which swallows the REST OF THE FILE:
+ *
+ *   input:  SELECT x$y$ FROM t;
+ *           DROP TABLE users;
+ *   guard:  destructive(0): (none)
+ *   split:  statements(1)   <- the whole file
+ *
+ * Measured on PostgreSQL 16.14, that file returns a column named `x$y$` and
+ * then drops the table, under an inventory that listed nothing. Collapsing a
+ * file to one statement also breaks the one-statement-per-round-trip contract
+ * `-- turbine:no-transaction` migrations depend on, which is this module's
+ * other production-destroying failure mode.
+ *
+ * Consuming the identifier run first IS Postgres's rule rather than an
+ * approximation of it, which is what keeps the converse working: a digit
+ * cannot START an identifier, so the `$$` in `SELECT 1$$x$$` still opens a
+ * body (PostgreSQL reports its syntax error at the string `$$x$$`, not at the
+ * number). Verified against the same server: `SELECT a$$b`, `SELECT$$x$$`, and
+ * `AS$$SELECT 1$$` are each ONE identifier token, and `naïve$col$` is a legal
+ * column name.
+ */
+const IDENT_START = /[A-Za-z_\u0080-\uFFFF]/;
+const IDENT_CONT = /[A-Za-z0-9_$\u0080-\uFFFF]/;
+/**
+ * A statement head that can carry a PG14+ SQL-standard routine body. Requiring
+ * it is what stops an ordinary transaction-control `BEGIN` from ever putting
+ * the walk into body mode, where semicolons stop terminating statements.
+ */
+const SQL_BODY_HEAD = /^\s*CREATE\s+(?:OR\s+REPLACE\s+)?(?:FUNCTION|PROCEDURE)\b/i;
+/**
+ * `BEGIN` as the immediately preceding token. The leading character class
+ * spells out what `\b` gets wrong here: `$` is an identifier character in
+ * Postgres, so `\bBEGIN` would also match the tail of `x$BEGIN`.
+ */
+const BEGIN_BEFORE_ATOMIC = /(?:^|[^A-Za-z0-9_$\u0080-\uFFFF])BEGIN\s+$/i;
+/**
+ * True when the identifier just scanned opens a PG14+ SQL-standard routine
+ * body: `CREATE FUNCTION ... BEGIN ATOMIC <stmt>; <stmt>; END`.
+ *
+ * Those inner semicolons are not statement terminators, and splitting there
+ * cost the guard the entire body. `CREATE FUNCTION purge() RETURNS void
+ * LANGUAGE SQL BEGIN ATOMIC DELETE FROM users; END;` became three fragments:
+ * the first headed `CREATE FUNCTION`, which matches no destructive rule, and
+ * the rest headless. A function whose only job is to empty a table therefore
+ * reported a clean inventory (verified live: the function creates, runs, and
+ * leaves zero rows). It also handed the no-transaction runner a fragment that
+ * cannot execute on its own.
+ *
+ * The two keywords must be ADJACENT (whitespace and comments aside) because
+ * both are unreserved. `CREATE FUNCTION f() RETURNS TABLE (begin int, atomic
+ * int)` is a legal header, verified on PostgreSQL 16, and reading that as a
+ * body would swallow every following statement up to the next `END`: the
+ * fail-OPEN direction, which is the one this guard cannot afford.
+ */
+function opensAtomicBody(word, before) {
+    return /^ATOMIC$/i.test(word) && BEGIN_BEFORE_ATOMIC.test(before) && SQL_BODY_HEAD.test(before);
+}
+/**
+ * True when the quote at `quoteAt` opens a Postgres escape string (`E'...'`),
+ * whose body treats a backslash as an escape character.
+ *
+ * The `E` must be a standalone token, so an identifier that merely ends in `e`
+ * does not turn the following literal into an E-string. Ordinary literals are
+ * left alone on purpose: with the modern `standard_conforming_strings = on`
+ * default, `'a\'` IS a complete string.
+ *
+ * Getting this wrong is not cosmetic. `E'p\'q'` is ONE literal; a tokenizer
+ * that ends it at the backslash-quote reads the rest of the file as string
+ * content, which merges following statements together (fatal for a
+ * `-- turbine:no-transaction` migration, which must issue one statement per
+ * round trip) and hides every later statement from the destructive guard.
+ */
+function isEscapeStringPrefix(sql, quoteAt) {
+    const prev = sql[quoteAt - 1];
+    if (prev !== 'E' && prev !== 'e')
+        return false;
+    const before = sql[quoteAt - 2];
+    return before === undefined || !/[A-Za-z0-9_$"]/.test(before);
+}
+/**
+ * Split a SQL script into its top-level statements.
+ *
+ * A semicolon inside a single-quoted string (including a backslash-escaping
+ * `E'...'` string), a double-quoted identifier, a dollar-quoted body, a PG14+
+ * `BEGIN ATOMIC` routine body, a line comment, or a block comment (which
+ * Postgres allows to NEST) does not terminate a statement. Nor does a `$` that
+ * merely continues an identifier open a dollar-quoted body, which is the same
+ * question asked from the other side (see IDENT_CONT and opensAtomicBody for
+ * what each of those two cost when they were wrong).
+ *
+ * Every returned statement is trimmed and carries no
+ * trailing semicolon; nothing is dropped, so a comment-only fragment comes back
+ * flagged rather than missing (the executor skips those, the scanner ignores
+ * them, and neither has to re-derive the fact).
+ */
+function tokenizeSql(sql) {
+    const out = [];
+    const n = sql.length;
+    let stripped = '';
+    let code = '';
+    let blocks = [];
+    let start = 0;
+    let i = 0;
+    // Inside a `BEGIN ATOMIC` routine body: 0 means no. The count tracks the
+    // `CASE ... END` expressions nested in it, which are the only other `END` a
+    // SQL-standard body can hold (`BEGIN ATOMIC SELECT CASE WHEN ... END; END`
+    // is legal, verified on PostgreSQL 16), so a body cannot end at the first
+    // `END` it happens to contain.
+    let atomicDepth = 0;
+    let atomicStart = 0;
+    const flush = (rawEnd) => {
+        const raw = sql.slice(start, rawEnd).trim();
+        const strippedTrimmed = stripped.trim();
+        if (raw.length > 0) {
+            out.push({
+                raw,
+                stripped: strippedTrimmed,
+                code: code.trim(),
+                blocks,
+                commentOnly: strippedTrimmed.length === 0,
+            });
+        }
+        stripped = '';
+        code = '';
+        blocks = [];
+    };
+    while (i < n) {
+        const ch = sql[i];
+        const next = sql[i + 1];
+        // Line comment: runs to the end of the line. The newline itself is left for
+        // the generic branch below, so line structure survives into `stripped`.
+        if (ch === '-' && next === '-') {
+            let j = i;
+            while (j < n && sql[j] !== '\n')
+                j++;
+            i = j;
+            code += ' ';
+            continue;
+        }
+        // Block comment. Postgres nests these, so depth is counted rather than
+        // scanning for the first `*\/`: `/* a /* b */ c */` is ONE comment, and a
+        // scanner that stops at the inner terminator treats ` c */ ...` as code.
+        if (ch === '/' && next === '*') {
+            let depth = 1;
+            let j = i + 2;
+            while (j < n && depth > 0) {
+                if (sql[j] === '/' && sql[j + 1] === '*') {
+                    depth++;
+                    j += 2;
+                }
+                else if (sql[j] === '*' && sql[j + 1] === '/') {
+                    depth--;
+                    j += 2;
+                }
+                else {
+                    j++;
+                }
+            }
+            i = j;
+            stripped += ' ';
+            code += ' ';
+            continue;
+        }
+        // Single-quoted literal. `''` always escapes a quote; inside an E-string a
+        // backslash escapes the next character too.
+        if (ch === "'") {
+            const escapes = isEscapeStringPrefix(sql, i);
+            let j = i + 1;
+            while (j < n) {
+                if (escapes && sql[j] === '\\') {
+                    j += 2;
+                    continue;
+                }
+                if (sql[j] === "'" && sql[j + 1] === "'") {
+                    j += 2;
+                    continue;
+                }
+                if (sql[j] === "'")
+                    break;
+                j++;
+            }
+            const end = Math.min(j + 1, n);
+            stripped += "''";
+            code += sql.slice(i, end);
+            i = end;
+            continue;
+        }
+        // Quoted identifier (`""` escapes a quote). Kept VERBATIM in `stripped`,
+        // because the destructive rules match on object names, but consumed as ONE
+        // token: an apostrophe inside a quoted name (`"customer's_orders"`) would
+        // otherwise open a string literal and hide every following statement, and a
+        // semicolon inside one (`"we;ird"`) would otherwise split a statement in
+        // half so that neither half matched anything.
+        if (ch === '"') {
+            let j = i + 1;
+            while (j < n) {
+                if (sql[j] === '"' && sql[j + 1] === '"') {
+                    j += 2;
+                    continue;
+                }
+                if (sql[j] === '"')
+                    break;
+                j++;
+            }
+            const end = Math.min(j + 1, n);
+            stripped += sql.slice(i, end);
+            code += sql.slice(i, end);
+            i = end;
+            continue;
+        }
+        // Identifier or keyword run, consumed as ONE token because Postgres does.
+        // This branch has to sit ABOVE the dollar-quote branch: a `$` that is part
+        // of an identifier is not an opener, and reading it as one swallows the
+        // rest of the file (see IDENT_CONT). It also gives the `BEGIN ATOMIC`
+        // tracking below a token boundary to work from, so `BEGINATOMIC` or a
+        // column named `atomic$` can never be mistaken for the keyword pair.
+        if (IDENT_START.test(ch)) {
+            let j = i + 1;
+            while (j < n && IDENT_CONT.test(sql[j]))
+                j++;
+            const word = sql.slice(i, j);
+            if (atomicDepth === 0) {
+                if (opensAtomicBody(word, stripped)) {
+                    atomicDepth = 1;
+                    atomicStart = j;
+                }
+            }
+            else if (/^CASE$/i.test(word)) {
+                atomicDepth++;
+            }
+            else if (/^END$/i.test(word) && --atomicDepth === 0) {
+                blocks.push(sql.slice(atomicStart, i));
+            }
+            stripped += word;
+            code += word;
+            i = j;
+            continue;
+        }
+        // Dollar-quoted body ($tag$ ... $tag$).
+        if (ch === '$') {
+            const tag = DOLLAR_TAG.exec(sql.slice(i))?.[0];
+            if (tag) {
+                const close = sql.indexOf(tag, i + tag.length);
+                const next = close === -1 ? n : close + tag.length;
+                blocks.push(sql.slice(i + tag.length, close === -1 ? n : close));
+                stripped += "''";
+                code += sql.slice(i, next);
+                i = next;
+                continue;
+            }
+        }
+        // Top-level statement terminator. The semicolons INSIDE a `BEGIN ATOMIC`
+        // routine body are not top level, so while one is open this falls through
+        // to the generic branch and the body stays one statement (see
+        // opensAtomicBody).
+        if (ch === ';' && atomicDepth === 0) {
+            flush(i);
+            i++;
+            start = i;
+            continue;
+        }
+        stripped += ch;
+        code += ch;
+        i++;
+    }
+    // A `BEGIN ATOMIC` body still open at EOF (a file that ends mid-routine, or
+    // an `END` this walk failed to find) is handed over anyway rather than
+    // dropped. Dropping it is the fail-OPEN direction: the body is executable
+    // SQL, and it would reach no scanner at all, whereas handing over a body
+    // that the server may reject as unterminated costs at most a confirmation
+    // prompt for a file that was never going to run.
+    if (atomicDepth > 0)
+        blocks.push(sql.slice(atomicStart));
+    flush(n);
+    return out;
+}
+/**
+ * Split a SQL script into individual executable statements on top-level
+ * semicolons, dropping comment-only fragments.
+ *
+ * This is the list `-- turbine:no-transaction` migrations issue one statement
+ * per `client.query()` call, which is the one production-destroying failure
+ * mode of the migration runner (a partial statement executed against
+ * production), so the behavior is pinned by exhaustive unit tests.
+ */
+function splitSqlStatements(sql) {
+    return tokenizeSql(sql)
+        .filter((s) => !s.commentOnly)
+        .map((s) => s.raw);
+}
+/**
+ * The whole script with comments removed, string and dollar-quoted literals
+ * emptied, and quoted identifiers preserved: the view the destructive rules are
+ * written against, joined back into one string.
+ *
+ * Statement-level consumers should prefer {@link tokenizeSql}, whose per-
+ * statement `stripped` is the same text without having to re-split it.
+ */
+function stripCommentsAndStrings(sql) {
+    return tokenizeSql(sql)
+        .map((s) => s.stripped)
+        .join('; ');
+}