Ink Bounded Floodfill

The bug this exists to prevent

A naive flood fill on a scanned floor plan leaks. Scans have hairline gaps where lines nearly meet, doorways are genuine openings, and a fill that escapes through either one returns the area of the entire floor instead of the room. Worse, it looks plausible — a big number, a valid polygon, nothing that reads as an error.

Closing the gaps by dilating the ink is the obvious fix, and it introduces the second bug: every boundary is now inset by the dilation radius, so every room measures small, consistently, by an amount nobody notices because all the rooms are wrong together.

The pipeline, and why each step is there

1. **Ink mask** — dark pixels become barriers.
2. **Dilate by half the gap tolerance** — closes hairline breaks and small door gaps so the fill
   cannot escape.
3. **Scanline flood** from the click point.
4. **Dilate the FILL back by the same radius** — undoes the inset from step 2, so the boundary
   lands on the true ink edge rather than inside it. This is the step that is easy to omit and
   produces uniformly undersized rooms.
5. **Moore boundary trace** to a polygon.

Steps 2 and 4 are a matched pair. Doing the first without the second is the silent-undersize bug; doing neither is the leak.

Drawn measurements act as dams

stampMeasurements lets a user-drawn linear close an opening the ink does not — a doorway you want treated as a wall for this measurement. Count dots deliberately do not dam anything: a point annotation is not a barrier, and treating it as one would silently split a room.

What this does NOT do

No PDF parsing, no rasterisation, no OCR, no scale detection. You supply the pixel data and the click point; it returns a polygon in base pixels. Converting to real-world units is yours.

It does not recognise rooms semantically — it fills a connected region. An open-plan space with no dividing ink is one region, correctly.

Verified

8 tests covering the fill, gap closing, boundary recovery after dilation, drawn linears damming an opening, and count dots not damming.

Not covered: no test runs against a real scanned drawing. Fixtures are synthetic rasters, so performance and behaviour on noisy scans — speckle, JPEG artefacts, faint pencil — is untested.