docs: refresh README with current pipeline + accurate license

The "How it works" steps now match the five-stage Upload → EXIF →
Datums → Deskew → Measure flow (the old README still described a
four-step pipeline ending at "Run correction"). The algorithm section
is rewritten to describe what the solver actually does today: primary-
gauge selection, the alternating-minimization loop around findHomography,
the per-type correspondence builders (rectangle/line/ellipse),
confidence-as-replication weighting, oscillation detection, output
bounds clamp, and per-datum residual reporting.

Adds the three example tiles from the upload page, fixes the presets
list to mention circles, and corrects the license to match the GPLv3
LICENSE file already in the repo.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

README.md

@@ -4,18 +4,50 @@ Client-side image deskewing tool. Upload a photo taken at an angle, place refere
 
 Everything runs in the browser -- no server, no uploads.
 
+## Example
+
+<table>
+  <tr>
+    <td width="50%" align="center">
+      <img src="public/example-before.jpg" alt="Before — angled photograph with a paper sheet as a known-size reference" /><br/>
+      <em>Before — angled shot</em>
+    </td>
+    <td width="50%" align="center">
+      <img src="public/example-after.jpg" alt="After — perspective-corrected top-down view" /><br/>
+      <em>After — corrected perspective</em>
+    </td>
+  </tr>
+  <tr>
+    <td colspan="2" align="center">
+      <img src="public/example-measured.jpg" alt="Measured — the corrected image with measurement annotations baked in" /><br/>
+      <em>Measured — annotations baked in</em>
+    </td>
+  </tr>
+</table>
+
 ## How it works
 
-1. **Upload** a JPG or HEIC image (HEIC is converted automatically)
+1. **Upload** a JPG or HEIC image (HEIC is converted automatically). Past uploads are cached and reopen straight into Measure with their datums and annotations restored.
-2. **Review EXIF** data -- camera, lens, focal length
+2. **Review EXIF** data -- camera, lens, focal length.
-3. **Place datums** on the image -- rectangles or lines with known real-world dimensions
+3. **Place datums** on the image -- rectangles, lines, or ellipses (circles) with known real-world dimensions. Each datum carries a 1--5 confidence score. One datum can be flagged as the world-axis reference to fix the output orientation.
-4. **Run correction** -- OpenCV.js computes a perspective transform and outputs a corrected image
+4. **Deskew** -- OpenCV.js computes a perspective transform and produces a corrected image at a configurable px/mm scale, with per-datum residuals and a global RMS error reported.
+5. **Measure** -- annotate the corrected image with lines, rectangles, ellipses, circles, and angles in real-world units. Export the bare PNG, or the image with measurements baked in (full resolution or current viewport).
 
 ### The algorithm
 
-The highest-confidence rectangle datum defines the initial perspective correction via `getPerspectiveTransform`. All other datums (rectangles and lines) are projected through that transform and measured. Per-axis weighted scale corrections are computed from the discrepancies, folded back into the destination rectangle, and a single clean `warpPerspective` produces the output. One matrix, one warp, no post-hoc distortion.
+The pipeline picks a **primary** datum that fixes the output gauge (orientation, position, scale): any datum the user explicitly flagged as the world-axis reference, otherwise the highest-confidence rectangle, then ellipse, then line. The primary's known dimensions are mapped onto an axis-aligned output frame via `cv.getPerspectiveTransform` -- that's the warm-start homography.
 
-Datum confidence scores (1--5) act as weights in the correction.
+That homography is then refined by an alternating-minimization loop around `cv.findHomography` (which internally runs DLT + Levenberg-Marquardt). On every outer pass, each datum is turned into output-space point correspondences using its shape constraint:
+
+- **Rectangle** -- Procrustes-fit an ideal `w × h` rectangle to the current projection of the user-marked corners.
+- **Line** -- preserve the projected midpoint and direction, rescale to the expected length.
+- **Ellipse** -- sample 12 points along the user ellipse, project them, then radially snap to a circle of the expected diameter centred on the projected user-marked centre. Forces circles to stay circular under perspective.
+
+Confidence scores (1--5) are realised as integer replication of correspondences -- `cv.findHomography` has no native weighting. The primary additionally gets a ×3 gauge boost so its anchors don't drift while secondary datums are being satisfied. The loop runs until the homography stops moving (max-entry relative delta below 1e-6) or 30 iterations, with period-2 oscillation detection logging a warning if the alternating minimization fails to converge.
+
+A single `cv.warpPerspective` with Lanczos resampling produces the corrected image at the requested px/mm scale; output bounds are derived by projecting the source image corners through the final homography and clamped to a 12288 px maximum dimension to keep WASM heap usage bounded.
+
+Per-datum residuals are reported alongside the result: edge length error and corner perpendicularity for rectangles, length error for lines, and isotropy / skew / equivalent-diameter error from the projected conic for ellipses, plus a global RMS percentage across all of them.
 
 ## Quick start
 
@@ -57,7 +89,7 @@ pnpm type-check # vue-tsc
 
 ## Datum presets
 
-Rectangles: A3, A4, A5, A6, 15x10 cm. Custom dimensions supported. Lines: any length.
+Rectangles: A3, A4, A5, A6, 15×10 cm. Circles: €1, €2, US 25¢, UK 1p, CD. Custom dimensions supported on every type; lines accept any length.
 
 ## How Skwik compares
 
@@ -73,8 +105,8 @@ There are plenty of tools that do *part* of what Skwik does, but none that combi
 | [Toolschimp Image Measure](https://www.toolschimp.com/image-measure) | ✅ | ❌ | ✅ | ✅ | ❌ |
 | [Aspose Deskew](https://products.aspose.app/imaging/image-deskew) | ❌ | ❌ | ❌ | ❌ | ❌ |
 
-Most deskew tools just pull 4 corners to a rectangle without any real-world dimensions -- the output has no scale. Most measurement tools calibrate against a single reference and don't correct perspective. Skwik uses multiple weighted datums (rectangles + lines, each with a confidence score) to solve both problems in one pass, and lets you measure distances or export with a scale bar on the corrected image.
+Most deskew tools just pull 4 corners to a rectangle without any real-world dimensions -- the output has no scale. Most measurement tools calibrate against a single reference and don't correct perspective. Skwik uses multiple weighted datums (rectangles, lines, and ellipses, each with a confidence score) to solve both problems in one pass, and lets you measure distances or export with a scale bar on the corrected image.
 
 ## License
 
-MIT
+GNU General Public License v3.0