feat(plot): import PDFs, HEIC and camera photos, identified by content

CHANGELOG.md

@@ -9,6 +9,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **Import from anything.** One import that identifies a file by its *contents* rather than
+  its name — operators rename files, and a phone hands over an `image.jpg` that is really
+  HEIC. **PDF** is now a first-class source: a vector page imports as lines at the size the
+  page declares, keeping its page position so registration works exactly as it does for an
+  SVG; a page that is really a scan goes to the conversion wizard; a multi-page file asks
+  which page. HEIC is named rather than just rejected ("convert it on your phone"), anything
+  unreadable says so instead of adding an empty artwork, and phones get a **Take photo**
+  button so a sketch on paper can be traced without leaving the app. pdf.js is loaded on
+  demand, so the bundle every client downloads to jog the machine is unchanged.
+
+### Fixed
+
+- The daemon served `.mjs` as `application/octet-stream`, which browsers refuse to execute as
+  a module — a lazily-loaded chunk would have worked in the dev server and failed only in the
+  packaged build. `.wasm` and `.map` were missing too.
 - **Holding tabs for cutting.** Short uncut bridges that keep each piece attached to the
   sheet until it is snapped out by hand — the opposite of the overcut, and wanted just as
   often. Count, width and a minimum contour length are set in the knife profile; bridges are

README.md

@@ -25,9 +25,14 @@ unattended **Raspberry Pi** setup needs.
 
 ## What it does
 
-- **One import, with a wizard for images.** SVGs are flattened to polylines in the browser
-  DOM (`getPointAtLength` / `getCTM`, no dependencies) and go straight onto the page at
-  their real size. A raster image opens the **import wizard**: adjust it (brightness,
+- **One import for every source.** What a file *is* is read from its contents, not its name:
+  **SVG** and **vector PDF** go straight onto the page at the real size they declare (keeping
+  their page position, so registration works); a **PDF that is really a scan**, a photo, or
+  any raster the browser can decode opens the conversion wizard; a **multi-page PDF** asks
+  which page; **HEIC** says what to do about it rather than failing silently; and on a phone
+  there is a **Take photo** button, so a sketch on paper can be photographed and traced at
+  the machine. SVGs are flattened to polylines in the browser DOM (`getPointAtLength` /
+  `getCTM`, no dependencies). A raster image opens the **import wizard**: adjust it (brightness,
   contrast, invert, rotate, crop), then choose how it becomes lines — **Outline**
   (marching-squares iso-contours), **Hatching** (parallel lines, denser where the image is
   darker), **Crosshatch** (more directions the darker the tone), **Stippling** (dots placed

gateway/server.ts

@@ -521,10 +521,16 @@ async function handleCommand(ws: WebSocket, msg: ClientMessage) {
 const MIME: Record<string, string> = {
   '.html': 'text/html',
   '.js': 'text/javascript',
+  // A lazily-loaded chunk (the PDF worker) is served as .mjs. Browsers refuse
+  // to execute a module served as application/octet-stream, so without this the
+  // feature fails only in the packaged build — never in the dev server.
+  '.mjs': 'text/javascript',
   '.css': 'text/css',
   '.svg': 'image/svg+xml',
   '.png': 'image/png',
   '.json': 'application/json',
+  '.map': 'application/json',
+  '.wasm': 'application/wasm',
 };
 const httpServer = createServer(async (req, res) => {
   try {

openspec/changes/archive/2026-09-12-import-sources/.openspec.yaml

@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-09-12

openspec/changes/archive/2026-09-12-import-sources/design.md

@@ -0,0 +1,43 @@
+## Context
+
+`onLoadImage` branched on `file.type` and the `.svg` extension, and everything else went to
+`imageToField`, which is `<img>` decoding — so whatever the browser could decode worked, and
+anything else produced an error with no explanation.
+
+## Decisions
+
+- **Sniff the bytes.** Extensions lie, MIME types are often just the extension in disguise, and the
+  cost of being wrong is a silent empty artwork rather than an error. The sniffer is pure, so each
+  format's signature is pinned by a test rather than discovered on a machine with vinyl in it. The
+  MIME type is kept as a tiebreak only where the bytes are genuinely ambiguous (a long XML prologue
+  pushing `<svg` past the sniffed window).
+- **HEIC gets its own answer.** It is an ISO-BMFF file like an MP4, distinguished by its brand.
+  Naming it means the app can say what to do about it.
+- **The pdf.js operator decoder is pure and the op codes are asserted at load.** Decoding runs on a
+  plain operator list, so it is testable without a PDF or a DOM; the hard-coded codes are checked
+  against the library's own at load time, because a renumbering in a future version would not break
+  the build — it would just make every PDF import as blank.
+- **A PDF page is flipped at the boundary.** PDF user space is y-up from the bottom-left. The flip
+  happens in the decoder, so everything downstream keeps the single convention it already assumes.
+- **Vector or scan is decided by what the page paints**, not by a setting: if there are paths, they
+  are what the operator wants; if there are none, the page is rasterised and the wizard takes over.
+- **Rasterising uses print intent**, although nothing is printed. pdf.js drives its display render
+  loop with `requestAnimationFrame`, which does not fire in a background tab — an import started
+  and then left alone would hang for ever. This is a rasterisation to trace, not something anyone
+  watches.
+- **The page is painted white first.** A PDF page has no background of its own; without it,
+  transparent areas read as black and the whole page traces as one solid blob.
+- **pdf.js is imported on demand.** It is by far the largest dependency here and most sessions
+  never open a PDF, so it must not sit in the bundle every client downloads to jog the machine.
+- **The camera is a separate control, mobile-only.** Putting `capture` on the main import would
+  force the camera and take away the file picker. The placement controls stay on the desktop
+  layout; what a phone gets is the ability to photograph a sketch and trace it.
+
+## Risks / Trade-offs
+
+- Vector PDFs vary enormously. This decoder handles paths, transforms and the save/restore stack;
+  it ignores clipping, patterns and shading, and counts text runs rather than tracing glyph
+  outlines (which would plot hollow letters). A page that is mostly text will import as very little
+  and say so.
+- An unknown path opcode stops that chunk rather than guessing: opcodes and coordinates share one
+  stream, so reading past one turns the rest of the path into nonsense geometry.

openspec/changes/archive/2026-09-12-import-sources/proposal.md

@@ -0,0 +1,34 @@
+## Why
+
+The import button said "+ Image" but meant "SVG, PNG, JPEG or WebP", and it decided which by looking
+at the file's *name*. Half an operator's sources are neither: the print-and-cut files come as PDF,
+phones produce HEIC, and a sketch on paper is a photograph that has to be taken, not found.
+
+## What Changes
+
+- **Files are identified by content** (`src/plot/sniff.ts`), not by extension: operators rename
+  files, and a phone hands over an `image.jpg` that is really HEIC.
+- **PDF** (`src/plot/pdf.ts`): a vector page becomes lines at the size the page declares — with the
+  page position kept, so Place on page and the registration wizard work as they do for an SVG. A
+  page that is really a scan goes to the import wizard. Multi-page files ask which page.
+- **HEIC is named, not just rejected**: most browsers cannot decode it, and "convert it on your
+  phone" is useful where "could not read that image" is not.
+- **Take photo** on a phone, via the file input's `capture` attribute — so a sketch can be
+  photographed and traced standing at the machine.
+- **Anything unreadable says so**, instead of adding a silent empty artwork.
+
+## Capabilities
+
+### Modified Capabilities
+- `image-import-wizard`: one import accepts vector, raster and PDF, identified by content.
+- `responsive-control`: importing a photograph is available on a phone (the camera is the point).
+- `pi-deployment`: the daemon serves `.mjs`, `.wasm` and `.map` with correct types.
+
+## Impact
+
+- **Code:** new `src/plot/sniff.ts` and `src/plot/pdf.ts` (+ tests); `src/ui/App.tsx`;
+  `gateway/server.ts`.
+- **Dependency:** `pdfjs-dist`, imported on demand. The main bundle is unchanged; the library and
+  its worker are separate chunks, fetched only when a PDF is opened.
+- **Fixed in passing:** the daemon's static server had no MIME type for `.mjs`, so a lazily-loaded
+  module chunk would have failed in the packaged build while working perfectly in the dev server.

openspec/changes/archive/2026-09-12-import-sources/specs/image-import-wizard/spec.md

@@ -0,0 +1,49 @@
+## ADDED Requirements
+
+### Requirement: Sources are identified by content
+
+The import SHALL decide what a file is from its contents rather than its name or declared type, and
+SHALL accept vector files, PDFs and any raster image the browser can decode.
+
+#### Scenario: A renamed file
+
+- **WHEN** a file's extension does not match its contents
+- **THEN** it is imported according to what it actually is
+
+#### Scenario: Something that is not an image at all
+
+- **WHEN** the file cannot be identified
+- **THEN** the operator is told, and no empty artwork is added
+
+#### Scenario: An image the browser cannot decode
+
+- **WHEN** the file is HEIC and the browser has no decoder for it
+- **THEN** the message says so and what to do about it, rather than reporting a generic failure
+
+### Requirement: PDF pages import as lines or as an image
+
+A PDF page that contains vector paths SHALL be imported as lines at the size the page declares,
+keeping its position on the page. A page without paths SHALL be treated as a picture and sent to
+the conversion wizard. A file with several pages SHALL ask which page to import.
+
+#### Scenario: A cut file
+
+- **WHEN** the operator imports a vector PDF
+- **THEN** its paths are added at the page's real size, positioned as the page positions them, so
+  registration against a matching print works as it does for an SVG
+
+#### Scenario: A scan
+
+- **WHEN** the page contains no paths
+- **THEN** it is rendered and the conversion wizard opens on it
+
+#### Scenario: Several pages
+
+- **WHEN** the PDF has more than one page
+- **THEN** the operator chooses which one, and the imported artwork says which page it came from
+
+#### Scenario: Text in a PDF
+
+- **WHEN** a page contains text
+- **THEN** it is not traced — a PDF's text is glyph outlines, which plot as hollow letters — and the
+  operator is told how much was skipped

openspec/changes/archive/2026-09-12-import-sources/specs/pi-deployment/spec.md

@@ -0,0 +1,12 @@
+## ADDED Requirements
+
+### Requirement: The served GUI's assets have correct content types
+
+The daemon SHALL serve every asset type the built GUI produces with a content type browsers accept,
+including JavaScript module files (`.mjs`), WebAssembly and source maps.
+
+#### Scenario: A lazily-loaded module chunk
+
+- **WHEN** the GUI loads a feature whose code is split into a `.mjs` chunk
+- **THEN** the daemon serves it as JavaScript and the browser executes it, rather than refusing it
+  for its content type

openspec/changes/archive/2026-09-12-import-sources/specs/responsive-control/spec.md

@@ -0,0 +1,17 @@
+## ADDED Requirements
+
+### Requirement: Photographing a sketch from a phone
+
+On a phone the app SHALL offer taking a photograph and importing an image, so a drawing on paper
+can be photographed and traced without leaving the app. The placement controls remain on the
+desktop layout.
+
+#### Scenario: Taking a photo
+
+- **WHEN** the operator uses the take-photo control on a phone
+- **THEN** the camera opens, and the picture taken goes into the conversion wizard
+
+#### Scenario: Not on the desktop layout
+
+- **WHEN** the app is used on a wide screen
+- **THEN** the camera control is not shown, because there is nothing to photograph with

openspec/changes/archive/2026-09-12-import-sources/tasks.md

@@ -0,0 +1,34 @@
+## 1. Identification
+
+- [x] 1.1 `src/plot/sniff.ts`: PDF, SVG, raster signatures, HEIC by ISO-BMFF brand, MIME tiebreak
+- [x] 1.2 Unit tests for each signature, HEIC vs MP4, SVG after a BOM/prologue, MIME never
+      overriding a signature, and unknown files staying unknown
+
+## 2. PDF
+
+- [x] 2.1 `src/plot/pdf.ts`: pure operator-list decoder (paths, transforms, save/restore, curves),
+      page size in mm, y-flip, text/image reporting, op-code assertion
+- [x] 2.2 Unit tests: page size, flip, closed contour, cubic and quadratic curves, transform
+      apply/restore/nesting, text counted, images reported, degenerate subpath dropped, unknown
+      opcode stops the chunk, several subpaths, missing page box
+- [x] 2.3 Lazy loader: vector geometry, raster rendering with print intent and a white background,
+      worker as a Vite-fingerprinted URL
+
+## 3. Import flow
+
+- [x] 3.1 Route by sniffed kind; multi-page picker; vector page keeps its page offset
+- [x] 3.2 HEIC message; unreadable files reported rather than silently empty
+- [x] 3.3 Mobile-only "Take photo" with `capture`
+
+## 4. Daemon
+
+- [x] 4.1 Serve `.mjs`, `.wasm` and `.map` with correct content types
+
+## 5. Docs, gate, verification
+
+- [x] 5.1 README, CHANGELOG
+- [x] 5.2 `mise run ci` green
+- [x] 5.3 Verified in the browser against the packaged build: a vector PDF imports at its real size
+      (88.2 × 211.7 mm) with the right page offset; a three-page PDF asks which page and imports
+      page 2; a filled path imports as its outline; a page whose only content is an image opens the
+      wizard; the camera input appears only at phone width

openspec/specs/image-import-wizard/spec.md

@@ -133,3 +133,51 @@ image's overall brightness.
 
 - **WHEN** the image has no sharp boundary
 - **THEN** no lines are produced, and the wizard says there is nothing to add
+
+### Requirement: Sources are identified by content
+
+The import SHALL decide what a file is from its contents rather than its name or declared type, and
+SHALL accept vector files, PDFs and any raster image the browser can decode.
+
+#### Scenario: A renamed file
+
+- **WHEN** a file's extension does not match its contents
+- **THEN** it is imported according to what it actually is
+
+#### Scenario: Something that is not an image at all
+
+- **WHEN** the file cannot be identified
+- **THEN** the operator is told, and no empty artwork is added
+
+#### Scenario: An image the browser cannot decode
+
+- **WHEN** the file is HEIC and the browser has no decoder for it
+- **THEN** the message says so and what to do about it, rather than reporting a generic failure
+
+### Requirement: PDF pages import as lines or as an image
+
+A PDF page that contains vector paths SHALL be imported as lines at the size the page declares,
+keeping its position on the page. A page without paths SHALL be treated as a picture and sent to
+the conversion wizard. A file with several pages SHALL ask which page to import.
+
+#### Scenario: A cut file
+
+- **WHEN** the operator imports a vector PDF
+- **THEN** its paths are added at the page's real size, positioned as the page positions them, so
+  registration against a matching print works as it does for an SVG
+
+#### Scenario: A scan
+
+- **WHEN** the page contains no paths
+- **THEN** it is rendered and the conversion wizard opens on it
+
+#### Scenario: Several pages
+
+- **WHEN** the PDF has more than one page
+- **THEN** the operator chooses which one, and the imported artwork says which page it came from
+
+#### Scenario: Text in a PDF
+
+- **WHEN** a page contains text
+- **THEN** it is not traced — a PDF's text is glyph outlines, which plot as hollow letters — and the
+  operator is told how much was skipped

openspec/specs/pi-deployment/spec.md

@@ -88,3 +88,13 @@ Setting up a Pi SHALL be done by installing a single versioned Debian package (`
 - **WHEN** the package is purged
 - **THEN** the service is stopped and disabled and the package-managed files are removed
 
+### Requirement: The served GUI's assets have correct content types
+
+The daemon SHALL serve every asset type the built GUI produces with a content type browsers accept,
+including JavaScript module files (`.mjs`), WebAssembly and source maps.
+
+#### Scenario: A lazily-loaded module chunk
+
+- **WHEN** the GUI loads a feature whose code is split into a `.mjs` chunk
+- **THEN** the daemon serves it as JavaScript and the browser executes it, rather than refusing it
+  for its content type

openspec/specs/responsive-control/spec.md

@@ -49,3 +49,18 @@ From the phone view the operator SHALL be able to control a plot running on the
 - **WHEN** a plot is running and the operator uses the phone view to pause, resume, or stop it
 - **THEN** the command takes effect on the daemon-driven plot and the view's live state and progress update accordingly
 
+### Requirement: Photographing a sketch from a phone
+
+On a phone the app SHALL offer taking a photograph and importing an image, so a drawing on paper
+can be photographed and traced without leaving the app. The placement controls remain on the
+desktop layout.
+
+#### Scenario: Taking a photo
+
+- **WHEN** the operator uses the take-photo control on a phone
+- **THEN** the camera opens, and the picture taken goes into the conversion wizard
+
+#### Scenario: Not on the desktop layout
+
+- **WHEN** the app is used on a wide screen
+- **THEN** the camera control is not shown, because there is nothing to photograph with