bantamkit-mcp 0.35.7 → 0.35.8

package/dist/memory/store.js

@@ -38,11 +38,13 @@
  * handed — the two branches Python takes when `_snapshot` is pinned have no reachable
  * caller here. Everything else about both functions is the reference's.
  */
+import { randomBytes } from 'node:crypto';
+import { writeFileSync } from 'node:fs';
 import { basename } from 'node:path';
 import { BantamError } from '../errors.js';
 import { formatFact, parseFrontmatter, pySplit, pyStrip, todayLocal } from './factfile.js';
 import { PyScalar } from './pyyaml.js';
-import { asPyOSError, cmpCodepoint, matchesMd, pyExists, pyJoin, pyLexists, pyMkdirParents, pyMtimeDate, pyName, pyParent, pyParents, pyReadText, pyReplace, pyScandirNames, pySuffix, pyUnlink, pyWithSuffix, pyWriteText, PyOSError, sortedPathNames, } from './pyfs.js';
+import { asPyOSError, cmpCodepoint, matchesMd, pyExists, pyIsDir, pyJoin, pyLexists, pyMkdirParents, pyMtimeDate, pyName, pyNewlineOut, pyParent, pyParents, pyReadText, pyReplace, pyScandirNames, pySuffix, pyUnlink, pyWriteText, PyOSError, PyUnicodeDecodeError, sortedPathNames, } from './pyfs.js';
 /**
  * The exact bytes written into a `.bantamkit/.gitignore` the first time such a directory is
  * brought into existence (user ruling 2026-09-15, `.shiftwork/notes-job51/P0-probes.md`).
@@ -162,6 +164,42 @@ export const DURABLE_TYPES = ['feedback', 'user'];
  */
 const NAME_RE = /^[a-z0-9][a-z0-9-]*\n?$/;
 const NAME_PATTERN = '^[a-z0-9][a-z0-9-]*$';
+/**
+ * THE LONGEST NAME A SAVE ACCEPTS (job65/#156): `NAME_MAX_CHARS` in the reference, the same
+ * number. The fact lands as `<name>.md` through `<name>.md.<pid>.<8 hex>.tmp` (#161), so the
+ * name leaves room for that suffix under the 255-byte file-name limit. WAS: no cap, and a
+ * 301-character name came back as a raw `[Errno 63] File name too long: '<absolute path>'`.
+ * `NAME_RE` admits ASCII only, so characters and bytes are one count.
+ */
+export const NAME_MAX_CHARS = 128;
+/**
+ * ONE LINE MEANS ONE LINE (job65/#155): the reference's `_NOT_ONE_LINE`, the same ranges —
+ * every C0/C1 control plus U+2028/U+2029. Refused in a saved `description`; in a fact file on
+ * disk each run of them is READ as one space (see `facts()`), so `index.md` and the hook's
+ * injected header can never gain a line nobody saved.
+ */
+const NOT_ONE_LINE = /[\u0000-\u001f\u007f-\u009f\u2028\u2029]+/g;
+/**
+ * A LONE SURROGATE IS NOT TEXT (job65/#160). JSON can carry `\ud800` on its own; no UTF-8
+ * file can hold it. This port used to write U+FFFD and say "saved"; the reference raised a
+ * `UnicodeEncodeError`. A JS string holds a pair as two code units, so a surrogate is lone
+ * when it is a high one not followed by a low one, or a low one not preceded by a high one.
+ */
+const LONE_SURROGATE = /[\ud800-\udbff](?![\udc00-\udfff])|(?<![\ud800-\udbff])[\udc00-\udfff]/;
+function codepoint(unit) {
+    return `U+${unit.charCodeAt(0).toString(16).toUpperCase().padStart(4, '0')}`;
+}
+/** `U+XXXX` of the first lone surrogate in `text`, or `null` when it is all text. */
+export function loneSurrogate(text) {
+    const found = LONE_SURROGATE.exec(text);
+    return found === null ? null : codepoint(found[0]);
+}
+function notOneLine(text) {
+    NOT_ONE_LINE.lastIndex = 0;
+    const found = NOT_ONE_LINE.exec(text);
+    NOT_ONE_LINE.lastIndex = 0;
+    return found === null ? null : codepoint(found[0]);
+}
 export const DUPLICATE_JACCARD = 0.5;
 // ---- the precision gate (roadmap #6) ---------------------------------------------
 //
@@ -276,6 +314,13 @@ const FACT_UNREACHABLE = "a fact that could not be stat'd is not a fact that is
     'this path';
 const ARCHIVE_DESTINATION_UNREACHABLE = "a destination that could not be stat'd is not a name that is already archived, " +
     'and nothing has moved: the fact is still in facts/';
+/**
+ * job65/#141, #123: the codes that say a directory EXISTS and this process may not write into
+ * it — `NOT_WRITABLE_ERRNOS` in the reference. EACCES is the chmod case everywhere, EPERM an
+ * immutable flag or a Windows ACL, EROFS a read-only mount. On the mkdir path it is only half
+ * the test (`ensureDirs` also needs the root to be there); see the reference's docstring.
+ */
+const NOT_WRITABLE_CODES = new Set(['EACCES', 'EPERM', 'EROFS']);
 export class MemoryValidationError extends BantamError {
 }
 export class MemoryBudgetExceeded extends BantamError {
@@ -632,6 +677,13 @@ export class MemoryStore {
      * the failing leaf, or that strerror, would be a divergence manufactured by mkdir's
      * internals. `this.root` is the path the caller named and both sides agree on it, which is
      * what lets one conformance case pin this line byte for byte.
+     *
+     * AMENDED job65/#141: THE SENTENCE ABOVE IS NOW THE "NOT THERE" ARM ONLY. A `chmod 000`
+     * store answered it too, about a directory that exists. The second arm is still an outcome:
+     * the mkdir failed with one of `NOT_WRITABLE_CODES` AND `this.root` is a directory that is
+     * there, and then the refusal is `memory store is not writable: <root>: <strerror>; the
+     * directory exists but cannot be written to, so nothing was written`. The cwd-`/` divergence
+     * measured above is about a root that is NOT there, so it never reaches this arm.
      */
     ensureDirs() {
         // Checked BEFORE the mkdir below, per user ruling #2: this is the one moment that can
@@ -651,12 +703,30 @@ export class MemoryStore {
         catch (e) {
             if (!(e instanceof PyOSError))
                 throw e;
+            if (NOT_WRITABLE_CODES.has(e.code) && this.present())
+                throw this.unwritable(e);
             throw new MemoryValidationError(`memory store could not be created: ${this.root}; the directory is not` +
                 ' there and this filesystem would not make it, so nothing was written');
         }
         if (bantamkitDir !== null)
             ensureBantamkitGitignore(bantamkitDir, !bantamkitDirExistedBefore);
     }
+    /** `_present`: is `this.root` a directory that is there? "Not allowed to look" is "there". */
+    present() {
+        try {
+            return pyIsDir(this.root);
+        }
+        catch (e) {
+            if (!(e instanceof PyOSError))
+                throw e;
+            return true;
+        }
+    }
+    /** `_unwritable`: one sentence for a store that exists and refused a write (#141). */
+    unwritable(error) {
+        return new MemoryValidationError(`memory store is not writable: ${this.root}: ${error.strerror}; the directory` +
+            ' exists but cannot be written to, so nothing was written');
+    }
     // ---- ops ----
     /**
      * Write a fact, or refuse it as a near-duplicate of one already here.
@@ -679,6 +749,24 @@ export class MemoryStore {
         if (pyStrip(description || '') === '') {
             throw new MemoryValidationError('description must be a non-empty line');
         }
+        if (name.length > NAME_MAX_CHARS) {
+            throw new MemoryValidationError(`invalid name: ${name.length} characters; must be at most ${NAME_MAX_CHARS}`);
+        }
+        const texts = [
+            ['description', description],
+            ['body', body],
+            ...links.map((link, i) => [`links[${i}]`, link]),
+        ];
+        for (const [label, value] of texts) {
+            const bad = typeof value === 'string' ? loneSurrogate(value) : null;
+            if (bad !== null) {
+                throw new MemoryValidationError(`${label} holds a lone surrogate (${bad}), which is not text and cannot be stored as UTF-8; nothing was saved`);
+            }
+        }
+        const broken = notOneLine(pyStrip(description));
+        if (broken !== null) {
+            throw new MemoryValidationError(`description must be one line; it holds the control character ${broken}`);
+        }
         const newTokens = tokens(`${name} ${description}`);
         let existing = null;
         for (const fact of this.facts()) {
@@ -703,7 +791,16 @@ export class MemoryStore {
         };
         const path = this.factPath(name);
         const existed = pyExists(path) ? pyReadText(path) : null;
-        this.writeFact(fact);
+        // #141, the 555 shape: every directory was already there, so the fact write is the first
+        // syscall the mode refuses. Only the "exists, not writable" codes become the sentence.
+        try {
+            this.writeFact(fact);
+        }
+        catch (e) {
+            if (e instanceof PyOSError && NOT_WRITABLE_CODES.has(e.code))
+                throw this.unwritable(e);
+            throw e;
+        }
         try {
             this.checkIndexBudget();
         }
@@ -745,13 +842,13 @@ export class MemoryStore {
      * reference's `not 0.0 <= min_ratio <= 1.0` refuses it. That is a real mutant (M7b) and
      * it is killed on both sides.
      */
-    recall(query, k = null, stamp = true, minRatio = RECALL_MIN_SCORE_RATIO) {
+    recall(query, k = null, stamp = true, minRatio = RECALL_MIN_SCORE_RATIO, skipped) {
         if (!(minRatio >= 0 && minRatio <= 1))
             throw new MemoryValidationError(MIN_RATIO_RANGE);
         const limit = k ?? this.k;
         const q = tokens(query);
         let scored = [];
-        for (const fact of this.facts()) {
+        for (const fact of this.facts(undefined, skipped)) {
             let score = 0;
             for (const t of tokens(`${pyText(fact.name)} ${pyText(fact.description)}`))
                 if (q.has(t))
@@ -798,8 +895,8 @@ export class MemoryStore {
      * order, which is the order `recall` breaks a tie in. `stamp` dates the hit through
      * `stamp()` exactly as `recall` dates its hits; a miss writes nothing.
      */
-    lookup(name, stamp = true) {
-        for (const fact of this.facts()) {
+    lookup(name, stamp = true, skipped) {
+        for (const fact of this.facts(undefined, skipped)) {
             if (typeof fact.name === 'string' && fact.name === name) {
                 if (stamp)
                     this.stamp(fact);
@@ -828,11 +925,19 @@ export class MemoryStore {
      * and the budget error to `lint: FAIL — …` on stderr with exit 1.
      */
     lint() {
-        for (const fact of this.facts()) {
+        const findings = [];
+        for (const fact of this.facts(findings)) {
             if (!VALID_TYPES.includes(pyText(fact.type))) {
                 throw new MemoryValidationError(`fact '${pyText(fact.name)}' has invalid type '${pyText(fact.type)}'`);
             }
         }
+        // The two hand-edit shapes every other op now TOLERATES (#121, #122) are named here,
+        // first one in file order, because lint is the one surface whose job is to say what an
+        // operator should fix: an unquoted timestamp that `facts` read as a string, and a file
+        // whose name is not the fact's. Neither is a reason for `recall` or `compact` to refuse
+        // the whole store, and both are a reason for `lint` to exit 1.
+        if (findings.length > 0)
+            throw new MemoryValidationError(findings[0]);
         this.checkIndexBudget();
     }
     /**
@@ -924,9 +1029,13 @@ export class MemoryStore {
         // name, instead of surfacing as a `FileNotFoundError` out of the final write. `exist_ok`
         // semantics make it free for every store that is already there.
         this.ensureDirs();
+        // Keyed by the FACT, not by `fact.name` (#122): two files can carry one name, and a map
+        // keyed by name collapsed them — `indexBefore` reported 88 bytes for a 134-byte index, and
+        // the second file's line was never subtracted. `byEviction` sorts a copy of this same
+        // array, so the objects are the keys on both sides of the loop.
         const sizes = new Map();
         for (const fact of facts) {
-            sizes.set(pyHashKey(fact.name), Buffer.byteLength(this.indexLine(fact), 'utf8'));
+            sizes.set(fact, Buffer.byteLength(this.indexLine(fact), 'utf8'));
         }
         const all = [...sizes.values()];
         if (reserve === null) {
@@ -952,9 +1061,14 @@ export class MemoryStore {
         for (const fact of this.byEviction(facts)) {
             if (size <= target)
                 break;
-            const path = this.factPath(fact.name);
+            // THE FILE THE FACT WAS READ FROM, not `facts/<name>.md` (#122). The two are the same
+            // path for every fact this store wrote; they part on a hand-edited file whose
+            // frontmatter names another fact, and there `factPath(fact.name)` moved the WRONG file
+            // — the newest fact, under the name of the stalest — or raised for a name no file
+            // carries. `lint` names the mismatch.
+            const path = fact.source ?? this.factPath(fact.name);
             pyReplace(path, pyJoin(this.root, 'archive', pyName(path)));
-            const bytes = sizes.get(pyHashKey(fact.name));
+            const bytes = sizes.get(fact);
             size -= bytes;
             archived.push({
                 name: fact.name,
@@ -1186,6 +1300,18 @@ export class MemoryStore {
      * bytes `dream` MEASURES and the bytes the store WRITES are already one function. That is
      * the code-shape difference the Python half had to manufacture, not a behaviour one.
      */
+    /**
+     * Every fact this store can read, and `{file, reason}` for each file it could not.
+     *
+     * THE DECLARED TOLERANT LISTING (job65/#119), `readable_facts` in the reference. One empty,
+     * truncated, non-UTF-8 or unreadable fact file costs that file, not the store: the hook's
+     * SessionStart index and native export read through this, as `Memory.recallOutcome` reads
+     * through `recall(..., skipped)`. A `facts/` that cannot be listed still throws.
+     */
+    readableFacts() {
+        const skipped = [];
+        return { facts: this.facts(undefined, skipped), skipped };
+    }
     internals() {
         return {
             facts: () => this.facts(),
@@ -1343,10 +1469,11 @@ export class MemoryStore {
     /**
      * Every fact on disk, parsed.
      *
-     * The `read_text` is OUTSIDE the try in the reference, so a fact file that is not UTF-8
-     * raises a bare `UnicodeDecodeError` rather than a `malformed fact file …` sentence. That
-     * ordering is reproduced: `pyReadText` raises `PyUnicodeDecodeError` with CPython's own
-     * message and it is not caught here.
+     * AMENDED job65/#159: the `read_text` is now INSIDE the reference's try, so a fact file that
+     * is not UTF-8 is `malformed fact file <name>: 'utf-8' codec can't decode …` — naming the
+     * file — instead of the bare decoder message. `pyReadText` raises `PyUnicodeDecodeError`
+     * with CPython's own text and it is wrapped here; a `PyOSError` from the read still escapes
+     * as itself, as in the reference (`OSError` is not in its `except`).
      *
      * RULING — THE MESSAGE TEXT FOR A MALFORMED FILE. `_facts` interpolates the exception into
      * `malformed fact file {name}: {e}`, and that reaches the model. Three of the four shapes
@@ -1361,12 +1488,66 @@ export class MemoryStore {
      * carried as a must-differ `ruling` case in the conformance suite so it is re-measured on
      * every run instead of asserted once and forgotten.
      */
-    facts() {
+    /**
+     * `_facts`: every fact on disk, in `factPaths` order, with two hand-edit shapes tolerated.
+     *
+     * AN UNQUOTED TIMESTAMP IS READ AS ITS `str()` (#121). `yaml.safe_load` hands the reference
+     * a `datetime.date` for `created: 2026-01-01` and a `datetime.datetime` for the longer
+     * forms, while every file the store writes carries the quoted string. `recall` and `save`
+     * never minded, but `compact` sorts `(last_recalled or created, name)` and the sort raised
+     * `'<' not supported between instances of 'datetime.date' and 'str'` the moment one
+     * hand-edited file sat beside a written one — the only remedy the budget refusal names,
+     * dead on the store that named it. The scalar's `text` IS `str(value)` on the reference
+     * (`isoformat()` for a date, `isoformat(' ')` for a datetime), so both runtimes read the
+     * same string and the next write quotes it. Only timestamps are touched: an `int` or `bool`
+     * in these fields is left exactly as the scalar cases in `suites/store.mjs` pin it.
+     *
+     * `findings`, when given, collects one sentence per shape `lint` should name — the coerced
+     * timestamp, and a file whose name is not the fact's (#122) — without throwing, because
+     * every other caller tolerates both. `lint` is the only caller that passes an array.
+     */
+    /**
+     * `skipped`, when given, makes ONE BAD FILE COST THAT FILE (job65/#119): a file that is
+     * malformed or cannot be read is pushed as `{file, reason}` (the sentence this would have
+     * thrown) and every other file is still read. Only `recall`, `lookup` and `readableFacts`
+     * pass it; a save, a compaction or a lint keeps the strict read.
+     */
+    facts(findings, skipped) {
         const out = [];
         for (const path of this.factPaths()) {
-            const text = pyReadText(path); // outside the try, exactly as in the reference
+            if (skipped === undefined) {
+                out.push(this.readFact(path, findings));
+                continue;
+            }
+            try {
+                out.push(this.readFact(path, findings));
+            }
+            catch (e) {
+                if (!(e instanceof BantamError || e instanceof PyOSError))
+                    throw e;
+                skipped.push({ file: basename(path), reason: e.message });
+            }
+        }
+        return out;
+    }
+    /** One fact file, parsed, or the `malformed fact file` sentence for it (see `facts`). */
+    readFact(path, findings) {
+        const timestampText = (value) => value instanceof PyScalar && (value.ord.kind === 'date' || value.ord.kind === 'datetime')
+            ? value.text
+            : value;
+        {
             const file = basename(path);
             const malformed = (reason) => new MemoryValidationError(`malformed fact file ${file}: ${reason}`);
+            let text;
+            try {
+                text = pyReadText(path);
+            }
+            catch (e) {
+                // #159: `UnicodeDecodeError` is a `ValueError`, which the reference's `except` names.
+                if (e instanceof PyUnicodeDecodeError)
+                    throw malformed(e.message);
+                throw e;
+            }
             const parts = pySplit(text, '---\n', 2);
             if (parts.length !== 3) {
                 throw malformed(`not enough values to unpack (expected 3, got ${parts.length})`);
@@ -1385,21 +1566,51 @@ export class MemoryStore {
                 if (!(key in meta))
                     throw malformed(`'${key}'`);
             }
-            const rawCreated = meta['created'];
-            out.push({
-                name: meta['name'],
-                description: meta['description'],
+            const name = meta['name'];
+            const stamps = {};
+            for (const key of ['created', 'last_recalled']) {
+                const raw = meta[key];
+                const value = timestampText(raw);
+                if (findings !== undefined && value !== raw) {
+                    findings.push(`fact '${pyText(name)}' in ${file}: ${key} is the unquoted timestamp ${value}, ` +
+                        `read as '${value}'; quote it`);
+                }
+                stamps[key] = value;
+            }
+            // `path.stem`, spelled the way `archived()` spells it.
+            const suffix = pySuffix(file);
+            const stem = suffix === '' ? file : file.slice(0, file.length - suffix.length);
+            if (findings !== undefined && stem !== pyText(name)) {
+                findings.push(`fact '${pyText(name)}' in ${file}: the file name does not match the fact name; ` +
+                    `rename the file to ${pyText(name)}.md or the fact to ${stem}`);
+            }
+            const rawCreated = stamps['created'];
+            // #155: a description that is not one line is READ as one — each run of `NOT_ONE_LINE`
+            // is one space — and `lint` names it, like the other hand-edit shapes.
+            let description = meta['description'];
+            const broken = typeof description === 'string' ? notOneLine(description) : null;
+            if (broken !== null) {
+                description = description.replace(NOT_ONE_LINE, ' ');
+                if (findings !== undefined) {
+                    findings.push(`fact '${pyText(name)}' in ${file}: the description is not one line ` +
+                        `(it holds the control character ${broken}), read as '${description}'; make it one line`);
+                }
+            }
+            return {
+                name,
+                description,
                 type: meta['type'],
                 body: pyStrip(parts[2]),
                 links: pyList(meta['links']),
-                last_recalled: meta['last_recalled'] ?? null,
+                last_recalled: stamps['last_recalled'] ?? null,
                 // `or`, not `??`: every FALSY value falls back to the mtime, and that is now more
                 // than the empty string — `created: 0`, `created: false` and `created: 0.0` are all
-                // falsy in Python too, while `created: 2026-08-23` is a truthy `date` and is kept.
+                // falsy in Python too, while `created: 2026-08-23` is a truthy `date` (now its text)
+                // and is kept.
                 created: pyTruthy(rawCreated) ? rawCreated : pyMtimeDate(path),
-            });
+                source: path,
+            };
         }
-        return out;
     }
     /**
      * Date the recall.
@@ -1411,14 +1622,51 @@ export class MemoryStore {
      */
     stamp(fact) {
         fact.last_recalled = this.today();
-        this.writeFact(fact);
+        // job65/#123: BEST-EFFORT. A read-only store (chmod 555, a read-only mount, a CI cache)
+        // used to fail the whole recall here, after the hits were computed. A `PyOSError` from the
+        // write now leaves the old date on disk and the recall answers; see the reference's
+        // `_stamp` for why it is silent.
+        try {
+            this.writeFact(fact);
+        }
+        catch (e) {
+            if (!(e instanceof PyOSError))
+                throw e;
+        }
     }
-    /** Write via `<name>.md.tmp` and `os.replace`, so a reader never sees a half-written fact. */
+    /**
+     * Write via a PER-WRITER temp and `os.replace`, so a reader never sees a half-written fact
+     * and two writers never share a temp (job65/#161).
+     *
+     * WAS `<name>.md.tmp`, one temp per FACT: two servers saving one name raced on it (the
+     * second `replace` found nothing — `[Errno 2] … 'shared-name.md.tmp'`), and a write that
+     * died half-way left the temp beside the facts. NOW the shape `clockOut` and the hook ledger
+     * use (J65-1, J65-4): `<name>.md.<pid>.<8 hex>.tmp`, created with `wx` (O_EXCL), renamed
+     * onto the fact, and unlinked on any failure before the error goes on up. `*.md` listings
+     * never see it.
+     */
     writeFact(fact) {
         const path = this.factPath(fact.name);
-        const tmp = pyWithSuffix(path, '.md.tmp');
-        pyWriteText(tmp, formatFact(fact));
-        pyReplace(tmp, path);
+        const tmp = `${path}.${process.pid}.${randomBytes(4).toString('hex')}.tmp`;
+        try {
+            try {
+                writeFileSync(tmp, pyNewlineOut(formatFact(fact)), { encoding: 'utf8', flag: 'wx' });
+            }
+            catch (e) {
+                throw asPyOSError(e, tmp, undefined, 'crt');
+            }
+            pyReplace(tmp, path);
+        }
+        catch (e) {
+            try {
+                pyUnlink(tmp);
+            }
+            catch (cleanup) {
+                if (!(cleanup instanceof PyOSError))
+                    throw cleanup;
+            }
+            throw e;
+        }
     }
     /**
      * Write `index.md` from the facts on disk. THE OP THE ORIGINAL DEFECT DESTROYED.

package/dist/npminstall.d.ts

@@ -45,6 +45,24 @@ export declare function shlexJoin(command: readonly string[]): string;
  * batch file without a shell. Arguments are quoted because `cmd.exe` gets a string.
  */
 export declare function runInstaller(command: readonly string[]): [number, string];
+/** What one spawned command came back with: both streams apart, the status, and any spawn error. */
+export interface Spawned {
+    readonly status: number | null;
+    readonly stdout: string;
+    readonly stderr: string;
+    readonly error: (Error & {
+        code?: string;
+    }) | null;
+}
+/**
+ * One argv, spawned and captured, with the Windows shim handling `runInstaller` established
+ * (J65-19 moved it here so `--update`'s version check could spawn npm through the same rule).
+ *
+ * `timeoutMs` is `spawnSync`'s own deadline: the child is killed when it passes, and the result
+ * carries an `error` whose `code` is `ETIMEDOUT` — the one fact `selfupdate.fetchIndex` turns
+ * into its timeout refusal. No deadline when it is omitted, which is what an installer gets.
+ */
+export declare function spawnCommand(command: readonly string[], timeoutMs?: number): Spawned;
 /**
  * -1 when the index is ahead, 0 when they agree, 1 when the installed version is ahead.
  *

package/dist/npminstall.js

@@ -76,17 +76,38 @@ export function shlexJoin(command) {
  * batch file without a shell. Arguments are quoted because `cmd.exe` gets a string.
  */
 export function runInstaller(command) {
-    const [program, ...args] = command;
-    const win = process.platform === 'win32';
-    const quoted = win ? args.map((a) => (/[\s"^&|<>]/.test(a) ? `"${a.replace(/"/g, '\\"')}"` : a)) : args;
-    const done = spawnSync(program ?? '', quoted, { encoding: 'utf8', shell: win });
+    const done = spawnCommand(command);
     if (done.error) {
         // A missing `npm` is not a stack trace: it is a command that failed, and `COMMAND_FAILED`
         // is the sentence that says so and hands back what went wrong. 127 is the shell's own
         // code for "command not found", so the number means something to the person reading it.
         return [127, done.error.message];
     }
-    return [done.status ?? 1, `${done.stdout ?? ''}${done.stderr ?? ''}`];
+    return [done.status ?? 1, `${done.stdout}${done.stderr}`];
+}
+/**
+ * One argv, spawned and captured, with the Windows shim handling `runInstaller` established
+ * (J65-19 moved it here so `--update`'s version check could spawn npm through the same rule).
+ *
+ * `timeoutMs` is `spawnSync`'s own deadline: the child is killed when it passes, and the result
+ * carries an `error` whose `code` is `ETIMEDOUT` — the one fact `selfupdate.fetchIndex` turns
+ * into its timeout refusal. No deadline when it is omitted, which is what an installer gets.
+ */
+export function spawnCommand(command, timeoutMs) {
+    const [program, ...args] = command;
+    const win = process.platform === 'win32';
+    const quoted = win ? args.map((a) => (/[\s"^&|<>]/.test(a) ? `"${a.replace(/"/g, '\\"')}"` : a)) : args;
+    const done = spawnSync(program ?? '', quoted, {
+        encoding: 'utf8',
+        shell: win,
+        ...(timeoutMs === undefined ? {} : { timeout: timeoutMs }),
+    });
+    return {
+        status: done.status,
+        stdout: done.stdout ?? '',
+        stderr: done.stderr ?? '',
+        error: done.error ?? null,
+    };
 }
 /**
  * A dotted version as something orderable, deterministically, without claiming PEP 440.

package/dist/pyjsonschema.d.ts

@@ -2,6 +2,24 @@ import { type PyValue } from './pyjson.js';
 export declare class PyJsonSchemaUnsupported extends Error {
     constructor(message: string);
 }
+/**
+ * A schema the validator will not resolve, answered — not raised. `contract.schemaError`
+ * returns the message as the feedback sentence, so `validate_json` says `{"valid": false,
+ * "feedback": …}` where it used to say `isError` (#125). The Python side prints the same
+ * sentence from `contract._unsupported_schema` and from the `referencing` exceptions.
+ */
+export declare class PyJsonSchemaRefusal extends Error {
+    constructor(message: string);
+}
+/** The drafts this port validates. Draft 3 is looked up so it can be refused by name. */
+export type Draft = 4 | 6 | 7 | 2019 | 2020;
+/**
+ * `urllib.parse.urlsplit(uri).geturl()` — `_utils.URIDict.normalize`, the function every
+ * `$schema` goes through before `_META_SCHEMAS` is consulted. 3.12 strips leading C0/space,
+ * removes every tab/CR/LF, lowercases the scheme, and `urlunsplit` drops an empty query and
+ * an empty fragment. Measured in `tools/conformance/suites/validate.mjs` (`dialect/…`).
+ */
+export declare function pyUrlNormalize(uri: string): string;
 /**
  * `jsonschema.exceptions.ValidationError`, narrowed to the fields the relevance heuristic
  * and `schema_error` read.
@@ -11,6 +29,9 @@ export declare class PyJsonSchemaUnsupported extends Error {
  * is what `schema_error` interpolates. Keeping the two apart is not pedantry: `best_match`
  * ranks context errors by their relative paths and then `schema_error` prints an absolute
  * one, and a port that conflated them would name the wrong field inside every `anyOf`.
+ *
+ * `draft` is the dialect the keyword ran under — `error._type_checker`, which
+ * `_matches_type` consults, and the one place the dialect reaches `best_match`.
  */
 export interface PyValidationError {
     message: string;
@@ -21,11 +42,69 @@ export interface PyValidationError {
     path: (string | number)[];
     context: PyValidationError[];
     parent: PyValidationError | null;
+    draft: Draft;
 }
 export declare function absolutePath(error: PyValidationError): (string | number)[];
 /** `jsonschema._utils.equal`: JSON equality that refuses `True == 1`. */
 export declare function pyEqual(a: PyValue, b: PyValue): boolean;
-declare function iterErrors(instance: PyValue, schema: PyValue | boolean): PyValidationError[];
+/** What a keyword needs beyond its own value: the root resource and the dialect in force. */
+interface Ctx {
+    root: PyValue;
+    draft: Draft;
+    /** The ROOT's dialect — anchors are found by the resource's specification, not the leaf's. */
+    rootDraft: Draft;
+}
+declare function iterErrors(instance: PyValue, schema: PyValue | boolean, ctx?: Ctx): PyValidationError[];
+/**
+ * `re.search(pattern, string)` — the reference's dialect, translated, bounded (#131, #132).
+ *
+ * THE DIALECT IS PYTHON'S. Until job65 this was `new RegExp(pattern, 'u')` with a silent
+ * fallback to a non-`u` RegExp, so the two runtimes disagreed on 5 of 5 of #131's probes:
+ * `\d`/`\w` are Unicode-wide in `re` and ASCII in JS, `$` matches before a final newline in
+ * `re`, `(?P<x>…)` refused here, and the fallback read `\Z` as a literal `Z`. JSON Schema
+ * recommends ECMA-262, but the product's contract is the reference's verdict, so the pattern
+ * is parsed the way `re._parser` parses it and re-emitted as a `v`-mode RegExp whose every
+ * construct means what it means in `re`:
+ *
+ *   - `\d` → `\p{Nd}`; `\w` → `[\p{L}\p{N}_]` (`str.isalnum() or "_"`); `\s` → the 29
+ *     codepoints `str.isspace()` names (JS's `\s` adds U+FEFF and drops U+001C-1F, U+0085);
+ *     `\b`/`\B` are look-arounds over that `\w`, and `\B` never matches the empty string
+ *     (`SRE_AT_UNI_NON_BOUNDARY`, CPython 3.11-3.13). Measured over every codepoint: the
+ *     three classes are equal except for the codepoints Unicode 16 assigned after the
+ *     reference's Unicode 15 (5,004 for `\w`, 80 for `\d`, 0 for `\s`).
+ *   - `.` → `[^\n]` (JS also stops at `\r`, U+2028, U+2029); `$` → `(?=\n?$)`; `\A` → `^`;
+ *     `\Z` → `$`; `(?P<n>…)` → a group; `{,n}` → `{0,n}`; a `{` that is not a repeat is a
+ *     literal; every literal is emitted as `\u{…}`, so no JS syntax rule can reach it.
+ *
+ * A pattern `re` itself refuses is a `PyReError` (what `re.error` is on the reference: the
+ * schema check raises it). Constructs with no faithful translation — inline flags,
+ * look-behind (fixed-width in `re`, any width in JS), backreferences (a JS backreference to
+ * a group that did not take part matches the empty string; `re`'s fails), atomic groups,
+ * possessive quantifiers, conditional groups and `\N{…}` (no Unicode name table here) — are
+ * parsed, so that a broken pattern is still an error, and then REFUSED on both runtimes with
+ * one sentence (`unsupportedPatternSentence`; `contract._pattern_unsupported` on the
+ * reference), never answered differently.
+ */
+export declare class PyReError extends Error {
+    constructor(message: string);
+}
+interface PyReTranslation {
+    /** The `v`-mode source, or `null` when the pattern uses a construct refused above. */
+    source: string | null;
+}
+/** `re._parser.parse(pattern)` + `fix_flags`, re-emitted for a `v`-mode RegExp. */
+export declare function translatePyPattern(pattern: string): PyReTranslation;
+/** The sentence both runtimes answer for a pattern outside the shared dialect (#131). */
+export declare function unsupportedPatternSentence(pattern: string): string;
+/** The milliseconds a document's patterns may spend, in total, before the validation is refused (#132). */
+export declare const PATTERN_BUDGET_MS = 500;
+/** The sentence both runtimes answer when a document's patterns run past the budget (#132). */
+export declare function patternBudgetSentence(pattern: string): string;
+/**
+ * At most `limit` codepoints of `text`, then `…(N more)` (#153). Codepoints, not UTF-16
+ * units, because `len()` on the reference counts codepoints.
+ */
+export declare function clipText(text: string, limit?: number): string;
 /**
  * `jsonschema.exceptions.best_match`.
  *
@@ -38,13 +117,15 @@ declare function iterErrors(instance: PyValue, schema: PyValue | boolean): PyVal
  * parent stands.
  */
 export declare function bestMatch(errors: PyValidationError[]): PyValidationError | null;
-export declare function checkSchema(schema: PyValue | boolean, where?: string): void;
+export declare function checkSchema(schema: PyValue | boolean, where?: string, draft?: Draft): void;
 /**
  * `jsonschema.validate(instance, schema)` reduced to what `contract.schema_error` needs:
  * the single error `best_match` chooses, or `null`.
  *
- * `check_schema` runs FIRST, as it does in the reference. See `checkSchema` for what that
- * means here and for the five wrong answers it replaced.
+ * In the reference's order: `validator_for` (the dialect, which raises on a non-string
+ * `$schema`), then the one-resource pre-scan `contract.py` runs before `validate`, then
+ * `check_schema`, then the instance. A `PyJsonSchemaRefusal` out of here is an ANSWER —
+ * `contract.schemaError` returns its sentence — and a `PyJsonSchemaUnsupported` is a raise.
  */
 export declare function validate(instance: PyValue, schema: PyValue): PyValidationError | null;
 export { iterErrors };