@1agh/maude 1.0.11 → 1.3.0

package/apps/studio/exporters/pdf.ts

@@ -42,6 +42,8 @@ import { readArtboardPrintProp } from '../canvas-edit.ts';
 import { computeMarksGeometry, MARK_STROKE_PT, requiredSlugPt } from '../print/marks.ts';
 import { CSS_DPI, getPaperPreset, mmToPt, resolveBleedMm, trimSizeMm } from '../print/units.ts';
 import { exportShimPath, runShim } from './_runtime.ts';
+import type { ExportDegradation } from './degraded.ts';
+import { outlinePdf } from './ghostscript.ts';
 import {
   canvasShellUrl,
   type ExportContext,
@@ -49,6 +51,7 @@ import {
   type ExportOptions,
   type ExportResult,
 } from './index.ts';
+import { analyzePdfFonts, describeFontProblem, unprintableFonts } from './pdf-fonts.ts';
 import { clampDpi } from './png.ts';
 import type { Target } from './scope.ts';
 
@@ -205,6 +208,121 @@ export function parsePageFit(raw: unknown): string | null {
   return getPaperPreset(raw) ? raw : null;
 }
 
+/**
+ * How the export should treat text (issue #116).
+ *
+ *  - `keep`    — today's behaviour. Text stays selectable and searchable; the
+ *                fonts are whatever Chromium emitted. A preflight NOTICE is
+ *                attached when some of those are unprintable.
+ *  - `embed`   — assert every font is a real embedded font program. Type 3 is
+ *                a hard failure, named — never a silent pass.
+ *  - `outline` — every glyph becomes a vector path (Ghostscript). Print-safe
+ *                by construction; text is no longer selectable and the file
+ *                can grow substantially.
+ */
+export type PdfTextMode = 'keep' | 'embed' | 'outline';
+
+/**
+ * `options.text` — TOP-LEVEL, not nested under `pdfPrint`, deliberately.
+ * `pdfPrint`'s presence is what switches the boxes/marks post-pass on, and
+ * text handling is orthogonal to that: a plain non-print PDF can carry Type 3
+ * fonts just as easily as a leaflet can.
+ *
+ * Degrades to `keep` on anything unrecognised rather than throwing, per this
+ * adapter's standing rule that a malformed options bag must never break an
+ * export (`ExportOptions` is free-form with no schema gate upstream).
+ */
+export function parsePdfText(raw: unknown): PdfTextMode {
+  return raw === 'embed' || raw === 'outline' ? raw : 'keep';
+}
+
+/**
+ * Classify the fonts in the bytes we are about to hand back.
+ *
+ * On the SAVED BYTES, not the in-memory `PDFDocument`, and that is load-bearing:
+ * pdf-lib DEFERS `embedPage` until flush, so before a save the `pageFit` path's
+ * page XObject is a dangling ref and a font walk finds nothing at all. Analysing
+ * the document would silently report "no fonts" for exactly the scale-to-paper
+ * exports print shops receive. `test/pdf-font-preflight.test.ts` pins both
+ * halves of that.
+ *
+ * Returns null when pdf-lib cannot read the bytes. A preflight is a warning
+ * mechanism; it must never become a new way for a successful export to fail.
+ */
+async function analyzeFinalBytes(bytes: Uint8Array) {
+  try {
+    const doc = await PDFDocument.load(bytes, { updateMetadata: false });
+    return analyzePdfFonts(doc);
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Apply `options.text` to the assembled bytes — the one place all three modes
+ * are decided, so the fast path and the post-pass path can never disagree about
+ * what `embed` means.
+ */
+export async function applyTextMode(
+  bytes: Uint8Array,
+  mode: PdfTextMode,
+  hooks?: ExportHooks
+): Promise<{ body: Uint8Array; degraded?: ExportDegradation }> {
+  if (mode === 'outline') {
+    const outlined = await outlinePdf(bytes, { signal: hooks?.signal });
+    // Ghostscript exited 0 — but "it ran" is not "it worked". Re-run the same
+    // preflight over its output: if a font survived, the file LOOKS converted
+    // and is not, which is the failure this whole feature exists to make
+    // impossible. Better to fail than to hand back a plausible lie.
+    const after = await analyzeFinalBytes(outlined);
+    if (after && after.length) {
+      throw new Error(
+        `Ghostscript reported success but ${after.length} font(s) survived outlining ` +
+          `(${after.map((f) => f.name).join(', ')}). The PDF was not converted to curves; ` +
+          'do not send it to print.'
+      );
+    }
+    assertOutlinedSizeOk(outlined.byteLength);
+    return { body: outlined };
+  }
+
+  const fonts = await analyzeFinalBytes(bytes);
+  if (!fonts) return { body: bytes };
+  const bad = unprintableFonts(fonts);
+  if (!bad.length) return { body: bytes };
+
+  const { reason, remedy } = describeFontProblem(bad);
+  if (mode === 'embed') {
+    throw new Error(`text=embed: ${reason} ${remedy}`);
+  }
+  return {
+    body: bytes,
+    degraded: { fontsNotEmbedded: bad.map((f) => f.name), reason, remedy },
+  };
+}
+
+/**
+ * Outlining trades file size for print safety — glyphs become path data, and a
+ * photo-heavy piece can grow by an order of magnitude. That is a legitimate
+ * print deliverable, not a bug, so this reuses the SAME generous ceiling
+ * `assertTotalSizeOk` applies to captures rather than a tighter one; it exists
+ * to catch a runaway, not to cap ordinary print output.
+ *
+ * Checked separately because `assertTotalSizeOk` runs on the CAPTURED files,
+ * before assembly and long before outlining — the pass that actually creates
+ * the growth was downstream of every existing guard.
+ */
+export function assertOutlinedSizeOk(byteLength: number): void {
+  if (byteLength <= MAX_TOTAL_OUTPUT_BYTES) return;
+  const mb = Math.round(byteLength / 1024 / 1024);
+  const maxMb = Math.round(MAX_TOTAL_OUTPUT_BYTES / 1024 / 1024);
+  throw new Error(
+    `Outlined PDF is too large (~${mb}MB, limit ~${maxMb}MB) — converting text to curves ` +
+      'grows the file. Try exporting fewer artboards per file, a lower DPI, or ' +
+      'text=embed (which verifies the fonts without converting them).'
+  );
+}
+
 /**
  * PDF's own device-scale resolver — deliberately NOT `png.ts`'s
  * `resolveDeviceScale`, which defaults an absent `dpi` to `clampScale(undefined)`
@@ -361,6 +479,7 @@ export async function run(
   const timeoutSec = (options.timeoutSec as number | undefined) ?? 12;
   const printOpts = parsePdfPrintOptions(options.pdfPrint);
   const pageFit = parsePageFit(options.pageFit);
+  const textMode = parsePdfText(options.text);
   // Dogfood follow-up — raster CONTENT on the artboard (a dropped photo, a
   // large-format piece authored at a fraction of its physical size) needs a
   // real deviceScaleFactor to embed at print density; the page itself stays
@@ -407,11 +526,25 @@ export async function run(
     // Fast path — no print/pageFit options given: pass Chromium's vector PDF
     // straight through, exactly as before this feature (zero pdf-lib
     // round-trip overhead for the common "just export a PDF" case).
+    //
+    // The font preflight (`applyTextMode`) DOES parse here, and that is a
+    // deliberate cost. This path is what `/design:export pdf` with no options
+    // takes — i.e. the most likely way someone exports a leaflet — and a PDF
+    // that a print shop will reject while saying nothing is precisely the bug
+    // being fixed. The parse is a tokenize + object-table build (raw streams
+    // are copied by length, never decoded), and it degrades to silence rather
+    // than an error on a PDF pdf-lib cannot read.
     if (!needsPostPass) {
       const [only] = written;
       if (written.length === 1 && only) {
         const bytes = new Uint8Array(readFileSync(only.path));
-        return { filename: `${baseSlug}.pdf`, contentType: 'application/pdf', body: bytes };
+        const finished = await applyTextMode(bytes, textMode, hooks);
+        return {
+          filename: `${baseSlug}.pdf`,
+          contentType: 'application/pdf',
+          body: finished.body,
+          ...(finished.degraded ? { degraded: finished.degraded } : {}),
+        };
       }
       const out = await PDFDocument.create();
       for (const w of written) {
@@ -420,7 +553,13 @@ export async function run(
         for (const page of pages) out.addPage(page);
       }
       const bytes = await out.save();
-      return { filename: `${baseSlug}.pdf`, contentType: 'application/pdf', body: bytes };
+      const finished = await applyTextMode(bytes, textMode, hooks);
+      return {
+        filename: `${baseSlug}.pdf`,
+        contentType: 'application/pdf',
+        body: finished.body,
+        ...(finished.degraded ? { degraded: finished.degraded } : {}),
+      };
     }
 
     // Post-pass path — every page goes through pdf-lib (single-page case now
@@ -473,7 +612,17 @@ export async function run(
       }
     }
     const bytes = await out.save();
-    return { filename: `${baseSlug}.pdf`, contentType: 'application/pdf', body: bytes };
+    // Text handling runs LAST, on the fully assembled document: the boxes are
+    // set and the vector crop/registration marks are drawn, so they are part of
+    // what Ghostscript preserves rather than something applied to an already
+    // outlined file (pdf-lib could not re-open and re-mark it as cleanly).
+    const finished = await applyTextMode(bytes, textMode, hooks);
+    return {
+      filename: `${baseSlug}.pdf`,
+      contentType: 'application/pdf',
+      body: finished.body,
+      ...(finished.degraded ? { degraded: finished.degraded } : {}),
+    };
   } finally {
     rmSync(tmp, { recursive: true, force: true });
   }

package/apps/studio/exporters/remote.ts

@@ -54,6 +54,41 @@ function safeArtifactName(raw: string | null, format: Format): string {
   return base.slice(0, 255);
 }
 
+/**
+ * Reduce whatever the render service put in `x-maude-degraded` to a value the
+ * rest of the system can hold, mirroring `safeArtifactName` above.
+ *
+ * The header was previously `JSON.parse`d and cast straight to
+ * `ExportDegradation` — the one field on this boundary that skipped the file's
+ * own "everything crossing back is untrusted" rule. It mattered less while the
+ * only consumer read a boolean; it matters now that a degradation can carry a
+ * `fontsNotEmbedded` ARRAY which `client/export-center.jsx` renders with
+ * `.join(', ')`. A hostile or simply mixed-version worker sending an object
+ * there turns the export panel into a TypeError, and unbounded `reason` /
+ * `remedy` strings land verbatim in the on-disk history ledger.
+ *
+ * Returns undefined for anything unusable — a malformed note must never cost a
+ * real artifact, which is the same reason the caller swallows parse errors.
+ */
+function safeDegradation(raw: unknown): ExportDegradation | undefined {
+  if (!raw || typeof raw !== 'object' || Array.isArray(raw)) return undefined;
+  const o = raw as Record<string, unknown>;
+  const text = (v: unknown, max: number): string | undefined =>
+    typeof v === 'string' && v ? v.slice(0, max) : undefined;
+
+  const reason = text(o.reason, 2000);
+  if (!reason) return undefined; // `reason` is the only required field
+  const fonts = Array.isArray(o.fontsNotEmbedded)
+    ? o.fontsNotEmbedded.flatMap((f) => text(f, 200) ?? []).slice(0, 50)
+    : undefined;
+  return {
+    reason,
+    ...(o.audioDropped === true ? { audioDropped: true } : {}),
+    ...(fonts?.length ? { fontsNotEmbedded: fonts } : {}),
+    ...(text(o.remedy, 2000) ? { remedy: text(o.remedy, 2000) } : {}),
+  };
+}
+
 /** Read a response body with a hard byte ceiling — `arrayBuffer()` has none. */
 async function readCapped(res: Response, max: number): Promise<Uint8Array> {
   const declared = Number(res.headers.get('content-length'));
@@ -274,7 +309,7 @@ export async function renderRemotely(args: {
       const raw = /^[A-Za-z0-9+/=]+$/.test(degradedHeader)
         ? Buffer.from(degradedHeader, 'base64').toString('utf8')
         : degradedHeader;
-      degraded = JSON.parse(raw) as ExportDegradation;
+      degraded = safeDegradation(JSON.parse(raw));
     } catch {
       /* a malformed degradation note must not fail a real artifact */
     }

package/apps/studio/fs-watch.ts

@@ -15,6 +15,34 @@ export interface FsWatch {
 
 const DEBOUNCE_MS = 50;
 
+/**
+ * Runtime-state DIRECTORIES with no `fs:any` subscriber — see the note at the
+ * skip site for why this is deliberately NOT the sync taxonomy.
+ *
+ * `_comments/`, `_canvas-state/` and `_state/` are POINTEDLY absent: the first
+ * drives comment collab, and the other two are small and per-user, so there is
+ * no volume argument for suppressing them and a real risk in guessing. Add to
+ * this list only after confirming no `bus.on('fs:any' | 'fs:json' | …)`
+ * subscriber acts on the path.
+ */
+const RUNTIME_DIRS = ['_history', '_chat', '_trash', '_smoke'] as const;
+
+/**
+ * True when `rel` is one of `RUNTIME_DIRS` or a path inside one, on either
+ * separator (`fs.watch` hands back platform-native separators on Windows).
+ *
+ * Matching the BARE directory name is load-bearing, not tidiness: macOS raises
+ * an event for the containing directory as well as the file, so a write to
+ * `_chat/c-1.jsonl` arrives as BOTH `_chat/c-1.jsonl` and a bare `_chat`.
+ * Suppressing only the prefixed form would leave every streamed ACP update
+ * still waking every `fs:any` subscriber — the wakeup this list exists to stop.
+ */
+function isUnderRuntimeDir(rel: string): boolean {
+  return RUNTIME_DIRS.some(
+    (d) => rel === d || rel.startsWith(`${d}/`) || rel.startsWith(`${d}${path.sep}`)
+  );
+}
+
 export function createFsWatch(ctx: Context): FsWatch {
   let watcher: ReturnType<typeof watch> | null = null;
   const seen = new Map<string, number>();
@@ -48,10 +76,26 @@ export function createFsWatch(ctx: Context): FsWatch {
       watcher = watch(ctx.paths.designRoot, { recursive: true }, (_event, filename) => {
         if (!filename) return;
         // Skip our own runtime artifacts.
+        //
+        // NOT `isRuntimeStateRel` (sync/file-membership.ts), even though this
+        // list looks like a narrower copy of it. That classifier answers "may
+        // this file travel to a peer", and it excludes `_comments/` — but
+        // `_comments/` IS needed on this bus: collab/index.ts matches
+        // `^_comments/(.+)\.json$` on `fs:any`, and git/watch.ts treats it as
+        // versionable. Routing this list through the sync taxonomy would
+        // silently kill comment collaboration. The two lists answer different
+        // questions and are deliberately not the same list.
+        //
+        // What IS added here (#119) are the high-volume runtime directories
+        // with no `fs:any` subscriber. `_chat/` matters most: the bridge
+        // appends to it on every streamed update, so each chunk of an agent
+        // turn was raising an `fs:any` that woke canvas-list-watch, the
+        // activity tracker and gitWatch (the last running git status against
+        // the project's whole repo) for a file none of them can act on.
         if (filename.startsWith('_server.json')) return;
         if (filename.startsWith('_active.json')) return;
         if (filename.startsWith('_sync.json')) return;
-        if (filename.startsWith(`_history${path.sep}`) || filename.startsWith('_history/')) return;
+        if (isUnderRuntimeDir(filename)) return;
         emit(filename);
       });
     } catch (err) {