@jasy/pdf 1.0.0-alpha.7 → 1.0.0-alpha.9

package/dist/edit/document.js

@@ -0,0 +1,642 @@
+import { Lexer } from "./lexer.js";
+import { DEFAULT_MAX_STREAM_SIZE, inflateBounded, PdfStreamTooLargeError, } from "../utils/inflate.js";
+import { get, isDict, isRef, isStream, isString, nameOf, numberOf, stringsIn, } from "./objects.js";
+import { StandardAes256 } from "../crypto/security-handler.js";
+import { StandardLegacy } from "../crypto/legacy-handler.js";
+const EMPTY = new Uint8Array(0);
+/** Thrown when a file cannot be opened because of its encryption - always saying which of the reasons it
+ *  is: no password, wrong password, or a revision we do not implement. */
+export class PdfEncryptedError extends Error {
+    constructor(message) {
+        super(`@jasy/pdf: ${message}`);
+    }
+}
+/** The raw bytes of a `/Encrypt` entry that must be a string (`/U`, `/UE`). */
+const stringBytes = (o) => isString(o) ? o.bytes : undefined;
+export class PdfDocument {
+    constructor(bytes) {
+        this.bytes = bytes;
+        this.xref = new Map();
+        this.cache = new Map();
+        /** Object streams already unpacked, by their object number. */
+        this.objStmCache = new Map();
+        /** Decrypted stream payloads, keyed by the stream's data offset (unique per stream). */
+        this.plainStreams = new Map();
+        /** True when the cross-reference index was unusable and had to be rebuilt by scanning. Surfaced so
+         *  a caller can report it - a silently repaired file is exactly the kind of guess we avoid. */
+        this.recovered = false;
+        this.trailer = { kind: "dict", map: new Map() };
+        /** Whether this file indexes itself with an xref STREAM rather than a classic table. An incremental
+         *  update has to follow suit: mixing the two leaves readers that understand only one kind unable to
+         *  walk the chain back. */
+        this.usesXrefStream = false;
+        /** Ceiling on what one stream may inflate to; see `inflate.ts`. */
+        this.maxStreamSize = DEFAULT_MAX_STREAM_SIZE;
+    }
+    static load(bytes, opts = {}) {
+        const doc = new PdfDocument(bytes);
+        if (opts.maxStreamSize !== undefined)
+            doc.maxStreamSize = opts.maxStreamSize;
+        try {
+            doc.readXrefChain();
+        }
+        catch {
+            doc.xref.clear();
+        }
+        // No index, or one that does not lead to a catalog: rebuild it by scanning for object headers.
+        // Damaged tables are common enough in the wild that refusing here would reject usable files.
+        if (doc.xref.size === 0 || doc.catalog === undefined)
+            doc.rebuildByScanning();
+        return doc;
+    }
+    /**
+     * Open a document, decrypting it when it is password-protected.
+     *
+     * Async because key derivation and AES are (WebCrypto), and because decryption happens ONCE here,
+     * eagerly, rather than at every read: that keeps `getObject`, `streamData` and the whole form layer
+     * above them synchronous, which is the same trade the writer makes (`structure.ts` builds the handler
+     * before the synchronous constructor).
+     */
+    static async open(bytes, opts = {}) {
+        const doc = PdfDocument.load(bytes, opts);
+        if (!doc.isEncrypted)
+            return doc;
+        doc.security = await doc.securityHandlerFor(opts.password);
+        await doc.decryptEverything();
+        return doc;
+    }
+    /**
+     * Build the handler for this file's `/Encrypt` dictionary, or say precisely why we cannot.
+     *
+     * Every refusal names its reason: no password given, wrong password, or an encryption revision we do
+     * not implement. "Cannot open this file" would be the unhelpful answer, and guessing would be worse.
+     */
+    async securityHandlerFor(password) {
+        const enc = this.resolve(this.trailer.map.get("Encrypt"));
+        const filter = nameOf(this.lookup(enc, "Filter"));
+        const v = numberOf(this.lookup(enc, "V")) ?? 0;
+        const r = numberOf(this.lookup(enc, "R")) ?? 0;
+        if (filter !== "Standard") {
+            throw new PdfEncryptedError(`this PDF uses the ${filter ?? "unknown"} security handler; jasy implements only the standard one`);
+        }
+        // An EMPTY user password is legal and common: a document restricted only by an owner password opens
+        // everywhere without prompting. So always try "" first and only complain about a MISSING password
+        // once that has failed - refusing up front would reject files nothing else refuses.
+        const pw = password ?? "";
+        const rejected = () => {
+            throw password === undefined
+                ? new PdfEncryptedError("this PDF is encrypted; pass its password to open it")
+                : new PdfEncryptedError("wrong password for this PDF");
+        };
+        // `recoverFileKey` / `StandardLegacy.open` say "wrong password" by name; anything else that comes out
+        // of them is structural (a malformed /U, /UE, /O) and must not be reported as a bad password.
+        const isWrongPassword = (e) => String(e?.message ?? e).includes("wrong password");
+        const incomplete = (what) => {
+            throw new PdfEncryptedError(`this PDF's /Encrypt dictionary is ${what}; it cannot be opened`);
+        };
+        // R5 and R6 are AES-256 with one file key; R2 to R4 derive a key per object and may use RC4.
+        if (v === 5 && (r === 5 || r === 6)) {
+            const u = stringBytes(this.lookup(enc, "U"));
+            const ue = stringBytes(this.lookup(enc, "UE"));
+            if (u === undefined || ue === undefined || u.length < 48)
+                incomplete("incomplete");
+            // /UE is unwrapped with AES-CBC and no padding, so a length that is not a whole number of blocks
+            // is a broken file, not a wrong password.
+            if (ue.length === 0 || ue.length % 16 !== 0)
+                incomplete("/UE is not a whole number of AES blocks");
+            try {
+                return await StandardAes256.forReading(pw, u, ue, r);
+            }
+            catch (e) {
+                if (isWrongPassword(e))
+                    rejected();
+                throw new PdfEncryptedError(`this PDF's encryption could not be unwrapped: ${String(e?.message ?? e)}`);
+            }
+        }
+        if (r >= 2 && r <= 4) {
+            const o = stringBytes(this.lookup(enc, "O"));
+            const u = stringBytes(this.lookup(enc, "U"));
+            if (o === undefined || u === undefined)
+                incomplete("incomplete");
+            const cipher = this.streamCipher(enc, v);
+            if (cipher.kind === "identity") {
+                throw new PdfEncryptedError("this PDF declares encryption but selects the Identity crypt filter, which jasy does not model");
+            }
+            const ids = this.trailer.map.get("ID");
+            const firstId = Array.isArray(ids) ? stringBytes(this.resolve(ids[0])) : undefined;
+            try {
+                return StandardLegacy.open(pw, {
+                    revision: r,
+                    keyBits: cipher.keyBits,
+                    o: o,
+                    u: u,
+                    permissions: numberOf(this.lookup(enc, "P")) ?? 0,
+                    id: firstId ?? EMPTY,
+                    encryptMetadata: this.lookup(enc, "EncryptMetadata") !== false,
+                    aes: cipher.kind === "aes",
+                });
+            }
+            catch (e) {
+                if (isWrongPassword(e))
+                    rejected();
+                throw new PdfEncryptedError(`this PDF's encryption could not be set up: ${String(e?.message ?? e)}`);
+            }
+        }
+        throw new PdfEncryptedError(`this PDF uses standard security V${v}/R${r}, which jasy does not implement`);
+    }
+    /**
+     * Which cipher and key length actually apply to the streams.
+     *
+     * For V4 the answer is NOT the top-level `/Length`: the document names a crypt filter in `/StmF` and
+     * that filter carries its own `/CFM` and `/Length` (in BYTES, where the top-level one is in bits).
+     * `/Identity` means the streams are not enciphered at all - decrypting them anyway would destroy them.
+     */
+    streamCipher(enc, v) {
+        const topBits = numberOf(this.lookup(enc, "Length")) ?? 40;
+        if (v < 4)
+            return { kind: "rc4", keyBits: topBits };
+        const stmf = nameOf(this.lookup(enc, "StmF")) ?? "Identity";
+        if (stmf === "Identity")
+            return { kind: "identity", keyBits: topBits };
+        const cf = this.lookup(this.lookup(enc, "CF"), stmf);
+        const cfm = nameOf(this.lookup(cf, "CFM"));
+        if (cfm === "None" || cfm === undefined)
+            return { kind: "identity", keyBits: topBits };
+        const cfBytes = numberOf(this.lookup(cf, "Length"));
+        return {
+            kind: cfm === "AESV2" ? "aes" : "rc4",
+            keyBits: cfBytes !== undefined ? cfBytes * 8 : topBits,
+        };
+    }
+    /**
+     * Decrypt every string and stream, once, in place.
+     *
+     * Three things are deliberately NOT decrypted, because they were never encrypted (ISO 32000-1 7.6.2):
+     * the `/Encrypt` dictionary's own strings, a cross-reference stream, and strings of objects that live
+     * inside an object stream - the object stream was enciphered as a whole, so its members come out in
+     * the clear already.
+     */
+    async decryptEverything() {
+        const security = this.security;
+        if (!security)
+            return;
+        const encryptNum = isRef(this.trailer.map.get("Encrypt"))
+            ? this.trailer.map.get("Encrypt").num
+            : undefined;
+        // `/EncryptMetadata false` means the XMP metadata stream was left in the CLEAR on purpose - the
+        // industry norm, so indexers can read a file they cannot open. Deciphering it anyway destroys it,
+        // and it is what made an accessible+encrypted document unopenable.
+        const metadataInClear = this.lookup(this.resolve(this.trailer.map.get("Encrypt")), "EncryptMetadata") === false;
+        // Streams first: an object stream has to be readable before its members can be served at all.
+        for (const [num, entry] of this.xref) {
+            if (entry.kind !== "offset" || num === encryptNum)
+                continue;
+            const obj = this.getObject(num);
+            if (obj === undefined || !isStream(obj))
+                continue;
+            const type = nameOf(get(obj, "Type"));
+            if (type === "XRef")
+                continue; // never encrypted - it holds the index itself
+            if (type === "Metadata" && metadataInClear)
+                continue;
+            // Keyed by the stream's DATA OFFSET, not its object number: `streamData` receives the stream
+            // object and has no number in hand.
+            this.plainStreams.set(obj.start, await security.decrypt(this.rawStream(obj), { num, gen: entry.gen }));
+        }
+        // Everything parsed so far came out of ENCIPHERED bytes: `load` already resolved the catalog to check
+        // it, and if that catalog lives in an object stream its members were inflated from ciphertext and
+        // cached as garbage. Drop both caches now that `plainStreams` can serve the real data. The xref index
+        // itself stays - it was never encrypted.
+        this.cache.clear();
+        this.objStmCache.clear();
+        // Then the strings, in objects that sit directly in the file.
+        for (const [num, entry] of this.xref) {
+            if (entry.kind !== "offset" || num === encryptNum)
+                continue;
+            const obj = this.getObject(num);
+            if (obj === undefined)
+                continue;
+            for (const s of stringsIn(obj)) {
+                try {
+                    s.bytes = await security.decrypt(s.bytes, { num, gen: entry.gen });
+                }
+                catch {
+                    // The password already validated against /U, so the key is right and this string simply is
+                    // not ciphertext - a producer left it in the clear. Say so instead of surfacing an AES error.
+                    throw new PdfEncryptedError(`object ${num} contains a string that is not encrypted, so this file does not follow its own ` +
+                        "/Encrypt declaration and cannot be read safely");
+                }
+            }
+        }
+    }
+    // -------------------------------------------------------------------------------------------
+    // Cross-reference index
+    // -------------------------------------------------------------------------------------------
+    /** The byte offset in the `startxref` trailer at the very end of the file. */
+    findStartXref() {
+        const tail = new TextDecoder("latin1").decode(this.bytes.subarray(Math.max(0, this.bytes.length - 2048)));
+        const at = tail.lastIndexOf("startxref");
+        if (at < 0)
+            return undefined;
+        const n = /startxref\s+(\d+)/.exec(tail.slice(at));
+        return n ? Number(n[1]) : undefined;
+    }
+    /** Read the index at `startxref`, then walk `/Prev` back through earlier revisions. Entries already
+     *  known win: the NEWEST revision is read first, and an older one must not overwrite it. */
+    readXrefChain() {
+        let offset = this.findStartXref();
+        const seen = new Set();
+        while (offset !== undefined && offset >= 0 && offset < this.bytes.length && !seen.has(offset)) {
+            seen.add(offset);
+            const dict = this.readXrefSection(offset);
+            if (!dict)
+                break;
+            if (this.trailer.map.size === 0)
+                this.trailer = dict;
+            // A hybrid-reference file keeps a second, stream-based index for the same revision.
+            const hybrid = numberOf(dict.map.get("XRefStm"));
+            if (hybrid !== undefined && !seen.has(hybrid)) {
+                seen.add(hybrid);
+                this.readXrefSection(hybrid);
+            }
+            offset = numberOf(dict.map.get("Prev"));
+        }
+    }
+    /** One index section, either format. Returns its trailer dictionary. */
+    readXrefSection(offset) {
+        const lx = new Lexer(this.bytes, offset);
+        if (lx.eatWord("xref"))
+            return this.readXrefTable(lx);
+        // Otherwise it must be an `N G obj` header introducing an xref STREAM.
+        lx.pos = offset;
+        lx.parse(); // object number
+        lx.parse(); // generation
+        if (!lx.eatWord("obj"))
+            return undefined;
+        const obj = lx.parse();
+        if (obj === undefined || !isStream(obj))
+            return undefined;
+        if (nameOf(get(obj, "Type")) !== "XRef")
+            return undefined;
+        this.usesXrefStream = true;
+        this.readXrefStream(obj);
+        return obj.dict;
+    }
+    /** The classic table: `xref`, then `start count` subsections of 20-byte entries, then `trailer`. */
+    readXrefTable(lx) {
+        for (;;) {
+            if (lx.eatWord("trailer")) {
+                const t = lx.parse();
+                return t !== undefined && isDict(t) ? t : undefined;
+            }
+            const start = lx.parse();
+            const count = lx.parse();
+            if (typeof start !== "number" || typeof count !== "number")
+                return undefined;
+            for (let i = 0; i < count; i++) {
+                const offset = lx.parse();
+                const gen = lx.parse();
+                lx.skip();
+                const type = String.fromCharCode(this.bytes[lx.pos]);
+                lx.pos++;
+                if (typeof offset !== "number" || typeof gen !== "number")
+                    return undefined;
+                // 'n' = in use, 'f' = free. A free slot is not an object.
+                if (type === "n")
+                    this.note(start + i, { kind: "offset", offset, gen });
+            }
+        }
+    }
+    /**
+     * An xref STREAM: fixed-width binary rows, widths in `/W`, object numbers from `/Index`. Row type 1
+     * is a plain byte offset, type 2 points into an object stream, type 0 is a free slot.
+     */
+    readXrefStream(stream) {
+        const data = this.streamData(stream);
+        const w = get(stream, "W")?.map((x) => numberOf(x) ?? 0) ?? [];
+        if (w.length < 3)
+            return;
+        const rowLen = w[0] + w[1] + w[2];
+        if (rowLen === 0)
+            return;
+        const size = numberOf(get(stream, "Size")) ?? 0;
+        const indexRaw = get(stream, "Index");
+        const index = Array.isArray(indexRaw) ? indexRaw.map((x) => numberOf(x) ?? 0) : [0, size]; // the default covers every object
+        // Read a big-endian field of `n` bytes; a width of 0 means "use the default" (type defaults to 1).
+        const field = (at, n, dflt) => {
+            if (n === 0)
+                return dflt;
+            let v = 0;
+            for (let i = 0; i < n; i++)
+                v = v * 256 + data[at + i];
+            return v;
+        };
+        let row = 0;
+        for (let s = 0; s + 1 < index.length; s += 2) {
+            const first = index[s];
+            const count = index[s + 1];
+            for (let i = 0; i < count; i++, row++) {
+                const at = row * rowLen;
+                if (at + rowLen > data.length)
+                    return;
+                const type = field(at, w[0], 1);
+                const f2 = field(at + w[0], w[1], 0);
+                const f3 = field(at + w[0] + w[1], w[2], 0);
+                if (type === 1)
+                    this.note(first + i, { kind: "offset", offset: f2, gen: f3 });
+                else if (type === 2)
+                    this.note(first + i, { kind: "inStream", stream: f2, index: f3 });
+            }
+        }
+    }
+    /** Record an entry unless a NEWER revision already claimed that object number. */
+    note(num, entry) {
+        if (!this.xref.has(num))
+            this.xref.set(num, entry);
+    }
+    /**
+     * Last resort: scan the whole file for `N G obj` headers and believe those. Used when the index is
+     * missing or does not lead anywhere - and flagged via `recovered`, never silently.
+     */
+    rebuildByScanning() {
+        this.recovered = true;
+        this.xref.clear();
+        this.cache.clear();
+        // Unpacked object streams belong to the index we just threw away; keeping them would serve members
+        // resolved against offsets that no longer apply.
+        this.objStmCache.clear();
+        const text = new TextDecoder("latin1").decode(this.bytes);
+        for (const m of text.matchAll(/(?:^|[\s>])(\d+)\s+(\d+)\s+obj\b/g)) {
+            const num = Number(m[1]);
+            // A later definition wins here: scanning walks the file forwards, and an incremental update
+            // appends its newer version further down.
+            this.xref.set(num, {
+                kind: "offset",
+                offset: (m.index ?? 0) + m[0].indexOf(m[1]),
+                gen: Number(m[2]),
+            });
+        }
+        // A scan only finds top-level `N G obj` headers, but in a modern file most objects - the catalog
+        // included - live INSIDE an object stream. Unpack every /ObjStm we found so they can be reached at
+        // all; without this, recovering a compressed file finds a handful of streams and no catalog.
+        // A snapshot, not the live keys: the loop ADDS entries to the same map.
+        const scanned = Array.from(this.xref.keys());
+        for (const num of scanned) {
+            const stm = this.getObject(num);
+            if (stm === undefined || !isStream(stm) || nameOf(get(stm, "Type")) !== "ObjStm")
+                continue;
+            const data = this.streamData(stm);
+            const head = new Lexer(data, 0);
+            const count = numberOf(get(stm, "N")) ?? 0;
+            for (let i = 0; i < count; i++) {
+                const on = head.parse();
+                const off = head.parse();
+                if (typeof on !== "number" || typeof off !== "number")
+                    break;
+                // A top-level definition found by the scan is the newer one and keeps precedence.
+                this.note(on, { kind: "inStream", stream: num, index: i });
+            }
+        }
+        if (this.trailer.map.size === 0 || this.trailer.map.get("Root") === undefined) {
+            const t = /trailer\s*<</g;
+            let last, found;
+            while ((last = t.exec(text)) !== null)
+                found = last.index;
+            if (found !== undefined) {
+                const lx = new Lexer(this.bytes, found);
+                lx.eatWord("trailer");
+                const d = lx.parse();
+                if (d !== undefined && isDict(d))
+                    this.trailer = d;
+            }
+            // A file with only xref streams has no `trailer` keyword; its catalog is found by scanning.
+            if (this.trailer.map.get("Root") === undefined) {
+                for (const [num] of this.xref) {
+                    const o = this.getObject(num);
+                    if (o !== undefined && nameOf(get(o, "Type")) === "Catalog") {
+                        this.trailer.map.set("Root", { kind: "ref", num, gen: 0 });
+                        break;
+                    }
+                }
+            }
+        }
+    }
+    // -------------------------------------------------------------------------------------------
+    // Objects
+    // -------------------------------------------------------------------------------------------
+    /** The object with that number, or `undefined` when the file does not have it. */
+    getObject(num) {
+        if (this.cache.has(num))
+            return this.cache.get(num);
+        this.cache.set(num, undefined); // guard against a reference cycle while we resolve
+        const entry = this.xref.get(num);
+        let value;
+        if (entry?.kind === "offset")
+            value = this.parseObjectAt(entry.offset, num);
+        else if (entry?.kind === "inStream")
+            value = this.objectFromStream(entry.stream, num);
+        this.cache.set(num, value);
+        return value;
+    }
+    /** Parse `N G obj … endobj` at a byte offset, checking that it really is object `num`. */
+    parseObjectAt(offset, num) {
+        if (offset <= 0 || offset >= this.bytes.length)
+            return undefined;
+        const lx = new Lexer(this.bytes, offset);
+        const n = lx.parse();
+        lx.parse(); // generation
+        if (n !== num || !lx.eatWord("obj"))
+            return undefined;
+        return lx.parse();
+    }
+    /** Unpack an object stream once, then serve its members from the cache. */
+    objectFromStream(streamNum, num) {
+        let members = this.objStmCache.get(streamNum);
+        if (!members) {
+            members = new Map();
+            this.objStmCache.set(streamNum, members);
+            const stm = this.getObject(streamNum);
+            if (stm !== undefined && isStream(stm) && nameOf(get(stm, "Type")) === "ObjStm") {
+                const data = this.streamData(stm);
+                const count = numberOf(get(stm, "N")) ?? 0;
+                const first = numberOf(get(stm, "First")) ?? 0;
+                // The stream opens with `count` pairs of "object number, offset relative to /First".
+                const head = new Lexer(data, 0);
+                const pairs = [];
+                for (let i = 0; i < count; i++) {
+                    const on = head.parse();
+                    const off = head.parse();
+                    if (typeof on !== "number" || typeof off !== "number")
+                        break;
+                    pairs.push([on, off]);
+                }
+                for (const [on, off] of pairs) {
+                    const o = new Lexer(data, first + off).parse();
+                    if (o !== undefined)
+                        members.set(on, o);
+                }
+            }
+        }
+        return members.get(num);
+    }
+    /** Follow indirect references until a direct object is reached. */
+    resolve(o) {
+        let seen = 0;
+        while (o !== undefined && isRef(o)) {
+            if (++seen > 64)
+                return undefined; // a cyclic file must not hang us
+            o = this.getObject(o.num);
+        }
+        return o;
+    }
+    /** A dictionary entry, with the reference already followed. */
+    lookup(o, key) {
+        return this.resolve(get(this.resolve(o), key));
+    }
+    /**
+     * The file is password-protected. Every string and stream in it is enciphered, so anything read out
+     * of it is ciphertext until a decryption path exists - which is why the form layer refuses rather
+     * than handing back nonsense field names.
+     */
+    get isEncrypted() {
+        return this.trailer.map.get("Encrypt") !== undefined;
+    }
+    /** True once the password has been accepted and the contents decrypted, so what is read is plaintext.
+     *  Always true for a document that was never encrypted. */
+    get isReadable() {
+        return !this.isEncrypted || this.security !== undefined;
+    }
+    /**
+     * Encipher a value that is about to be written back INTO this document, or `undefined` when the
+     * document is not encrypted and the value goes in as it is.
+     *
+     * Needed because we decrypt everything on open: a string we then hand back to the writer is plaintext
+     * in memory, and writing it into an encrypted file unchanged would both leak it and produce a document
+     * that contradicts its own /Encrypt declaration.
+     */
+    async encryptForWrite(data) {
+        return this.security ? this.security.encrypt(data) : undefined;
+    }
+    /**
+     * Whether values written back into this document can be enciphered again.
+     *
+     * False for a legacy-encrypted file: we can READ RC4 and AES-128, but writing one would downgrade the
+     * user's document to a broken cipher. Asked BEFORE any work is done - otherwise the refusal depends on
+     * a string happening to be present in whatever we rewrite, which is a guarantee by accident.
+     */
+    get canReEncrypt() {
+        return !this.isEncrypted || (this.security?.canEncrypt ?? false);
+    }
+    /** The generation of an object as the file records it. Members of an object stream are always 0. */
+    generationOf(num) {
+        const entry = this.xref.get(num);
+        return entry?.kind === "offset" ? entry.gen : 0;
+    }
+    get catalog() {
+        const root = this.resolve(this.trailer.map.get("Root"));
+        return root !== undefined && isDict(root) ? root : undefined;
+    }
+    // -------------------------------------------------------------------------------------------
+    // Stream data
+    // -------------------------------------------------------------------------------------------
+    /** A stream's bytes as they sit in the file: sliced by `/Length`, still filtered and still encrypted. */
+    rawStream(stream) {
+        const length = numberOf(this.resolve(get(stream, "Length")));
+        if (length !== undefined && stream.start + length <= this.bytes.length) {
+            return this.bytes.subarray(stream.start, stream.start + length);
+        }
+        // A missing or wrong /Length happens; fall back to the next `endstream`.
+        const text = new TextDecoder("latin1").decode(this.bytes);
+        const end = text.indexOf("endstream", stream.start);
+        return this.bytes.subarray(stream.start, end < 0 ? this.bytes.length : end);
+    }
+    /** A stream's DECODED bytes: deciphered if the document is encrypted, then run through its filters. */
+    streamData(stream) {
+        // A decrypted payload replaces the bytes from the file; filters then run on the plaintext, which is
+        // the order the format prescribes - encryption wraps the already-filtered data.
+        const raw = this.plainStreams.get(stream.start) ?? this.rawStream(stream);
+        const filterRaw = this.resolve(get(stream, "Filter"));
+        const filters = (Array.isArray(filterRaw) ? filterRaw : [filterRaw])
+            .map((f) => nameOf(this.resolve(f)))
+            .filter((f) => f !== undefined);
+        const parmsRaw = this.resolve(get(stream, "DecodeParms"));
+        const parms = Array.isArray(parmsRaw) ? parmsRaw : [parmsRaw];
+        let data = raw;
+        filters.forEach((filter, i) => {
+            if (filter === "FlateDecode") {
+                try {
+                    // Per pass, because /Filter is a list and [/FlateDecode /FlateDecode] is a legal nested bomb.
+                    data = inflateBounded(data, this.maxStreamSize);
+                }
+                catch (e) {
+                    // A stream that blew the ceiling is an attack, not a stream we failed to decode - swallowing
+                    // it here would hand back silent garbage.
+                    if (e instanceof PdfStreamTooLargeError)
+                        throw e;
+                    return; // not our data to repair; hand back what we have
+                }
+                data = this.undoPredictor(data, this.resolve(parms[i]));
+            }
+            // Other filters (LZW, DCT, …) are left encoded on purpose: this reader only needs the structural
+            // streams - the index, object streams and appearances - and every one of those is Flate.
+        });
+        return data;
+    }
+    /**
+     * Undo a PNG predictor (`/Predictor` 10-15), which xref streams almost always use - the IRS form in
+     * our fixtures does. Each row is prefixed with its filter type and is decoded against the row above.
+     */
+    undoPredictor(data, parms) {
+        const predictor = numberOf(get(parms, "Predictor")) ?? 1;
+        if (predictor < 10)
+            return data; // 1 = none; 2 (TIFF) does not occur in the streams we read
+        const colors = numberOf(get(parms, "Colors")) ?? 1;
+        const bpc = numberOf(get(parms, "BitsPerComponent")) ?? 8;
+        const columns = numberOf(get(parms, "Columns")) ?? 1;
+        const bpp = Math.max(1, Math.ceil((colors * bpc) / 8));
+        const rowLen = Math.ceil((colors * bpc * columns) / 8);
+        const rows = Math.floor(data.length / (rowLen + 1));
+        const out = new Uint8Array(rows * rowLen);
+        let prev = new Uint8Array(rowLen);
+        for (let r = 0; r < rows; r++) {
+            const type = data[r * (rowLen + 1)];
+            const src = data.subarray(r * (rowLen + 1) + 1, (r + 1) * (rowLen + 1));
+            const row = out.subarray(r * rowLen, (r + 1) * rowLen);
+            for (let i = 0; i < rowLen; i++) {
+                const a = i >= bpp ? row[i - bpp] : 0; // left
+                const b = prev[i]; // up
+                const c = i >= bpp ? prev[i - bpp] : 0; // upper left
+                const x = src[i];
+                switch (type) {
+                    case 0:
+                        row[i] = x;
+                        break;
+                    case 1:
+                        row[i] = (x + a) & 0xff;
+                        break;
+                    case 2:
+                        row[i] = (x + b) & 0xff;
+                        break;
+                    case 3:
+                        row[i] = (x + ((a + b) >> 1)) & 0xff;
+                        break;
+                    case 4: {
+                        const p = a + b - c;
+                        const pa = Math.abs(p - a), pb = Math.abs(p - b), pc = Math.abs(p - c);
+                        row[i] = (x + (pa <= pb && pa <= pc ? a : pb <= pc ? b : c)) & 0xff;
+                        break;
+                    }
+                    default:
+                        row[i] = x;
+                }
+            }
+            prev = row;
+        }
+        return out;
+    }
+    /** Every object number the index knows, ascending - for walking a whole document. */
+    objectNumbers() {
+        return [...this.xref.keys()].sort((a, b) => a - b);
+    }
+}
+export { EMPTY };