Recover continent membership and territory shapes by segmenting map images

[jeremymanning]

Blocks the continent-bonus feature family in #1. Split out because it is a self-contained
computer-vision job, not part of the HTML-parsing path.

The problem

D12's markup exposes territory id, name, adjacency, and label coordinates — but not continent
membership. Verified against a real game page (/game/1889, World Classic): 42 territory
anchors, zero occurrences of data-region, region, or continent.

<a href="#5" id="territory-5" class="js_territory" data-territory="5"
   data-adjacencies="7,11,66,69" data-x="92" data-y="68" data-name="Northwest Territory">

The catalog reports num_regions per map (6 for World Classic, up to 36 for Land of Rus'), so
the count is known — the per-territory assignment is not. Continent bonuses are central to Risk
strategy, so the bonus-weighted-income feature in #1's position-strength model is blocked without
this.

Approach

Continents are colour-coded on the map artwork itself. Drive the map selection panel with
Playwright, capture each map's image, and segment it.

• /maps is robots-allowed, as are /assets/img/maps/<id>.large.jpg and
  /image/map/<id>.large.circles.jpg. The .circles variant overlays territory markers.
• Topology, by contrast, only exists on /game/<id>, which robots.txt disallows — so the image
  route is the better-founded path, not merely a fallback.
• This runs once, over 77 maps, not on any repeating schedule.

Segmentation should yield more than continents:

1. Continent membership — cluster territories by the artwork's fill colour at each known
   data-x/data-y, or by segmenting regions directly.
2. Territory polygons — actual boundaries rather than just label points. This is what basic approach #1 asks
   for when it says "high quality vector (plottable at any resolution) with labeled territories
   and connections"; the current renderer draws nodes at label coordinates, which is a weaker
   substitute.
3. Adjacency, as a cross-check — polygons that share a boundary are neighbours.

Validation is available, and should gate this

tests/fixtures/game_map1_territories.html holds ground truth for World Classic straight from
D12: 42 territories, 83 undirected edges, exact adjacency, real names. Segmentation-derived
adjacency must be checked against it rather than trusted. If the pipeline cannot reproduce those
83 edges, it is not ready to run on the other 76 maps.

Report per-map agreement rather than a single aggregate — a method that works on World Classic and
fails on dense 150-territory maps needs to be known as such, not averaged into looking fine.

Notes

• Map artwork is D12's and trademarked. Derived data (adjacency, membership, polygons) is factual
  and fine to publish; the source JPEGs are not to be redistributed, and figures must render
  our own geometry rather than reproducing their images.
• region_id already defaults to 0 throughout the codebase, and check_invariants treats a
  wholly region-less map as valid rather than an error (map 77, "Brecourt Manor", genuinely has 0
  regions). So this issue is additive — nothing needs restructuring to accept its output.

Depends on

#1 for the MapTopology model and the graph layer.

[jeremymanning]

Confirmed the image URL pattern by fetching: https://dominating12.com/assets/img/maps/<id>.large.jpg returns image/jpeg for every catalogued map, and imageUrl in the catalog matches that pattern for all 77 with no deviations. No authentication needed, and the path is robots-allowed — so image collection needs no Playwright and no session. A plain rate-limited fetch over the catalog's ids suffices; riskdyn pull-images (Task 11) already does this.

Do not iterate a numeric range. Catalogued map ids span 1-104 with 27 gaps:

3, 4, 9, 13, 14, 16, 17, 18, 19, 21, 23, 24, 26, 31, 37, 43, 45, 46, 47, 51, 55, 67, 81, 96, 97, 99, 101

A 1..77 loop would both miss real maps (ids 78-104) and hit uncatalogued ones.

The gaps do not 404 — map 3 serves a valid 264 KB JPEG. They appear to be retired or unlisted maps whose artwork is still hosted, but they are unusable here: with no catalog entry there is no num_regions to validate segmentation against, no num_territories, and no dimensions. Iterate the catalog's id list, not a range.

Images are ~240-265 KB each, ~20 MB for all 77. They are D12's artwork: cache outside the repo (data/raw/ is gitignored), do not commit or republish. Only derived geometry goes in figures.

[jeremymanning]

All 77 map images are now downloaded via riskdyn pull-images (robots-allowed /assets/ path,
rate-limited, ~27.5 MB, cached under the gitignored data/raw/map_images/). Three findings that
matter for segmentation:

1. One file's extension lies. Map 30 ("Tor", 42 territories) is served as a PNG with an alpha
channel despite the .jpg URL and filename:

	count
JPEG / RGB	76
PNG / RGBA	1 (map 30)

Any segmentation code must sniff the actual format rather than trust the extension, and must handle
a 4-channel image. Pillow sniffs automatically, but colour-based clustering that assumes 3 channels
will break on this one, and the alpha channel may carry transparency that reads as a spurious
region. Convert to RGB on load with an explicit background composite rather than dropping alpha.

2. Image dimensions match the catalog exactly, for all 77. Every image's (width, height)
equals the catalog's width/height fields. That is a free integrity check: assert it on load, and
a mismatch means either a stale catalog or a truncated download.

3. Territory label coordinates are in image pixel space. data-x/data-y from the topology
markup index directly into these images (World Classic: 1021x689 in both). So for any map where we
have topology, sampling the artwork at a territory's (x, y) gives that territory's fill colour
directly — no registration step needed. That is the cheapest possible path to continent membership
where topology exists.

The catch remains that topology exists for exactly one map (World Classic, from a saved /game/
page) until data-use permission arrives. For the other 76 the pipeline has to recover territory
regions from the image alone, which is the harder job described above.

[jeremymanning]

Stage 1 (geometry) — implemented, red-teamed, reworked. Not done.

riskdyn/segment/ using Segment Anything (facebook/sam-vit-huge, pinned revision, MPS,
~30-40 s/map). Suite 121 → 133. Commits 1288dbc, b7cdd08, 4d88226, 3523dca, then
ded5ea4, e0293a7, db2ad56.

SAM was the right call: it is class-agnostic, so it copes with the catalog's real variety
(saturation 26-218, colour counts 47-202, edge density 11-37) where any colour-threshold pipeline
tuned on World Classic would have died on Arctic Circle's near-white territories.

What works

The territory outlines themselves are accurate. Median mask-vs-polygon IoU 0.984, only 0.30%
of pixels double-covered. Determinism verified across separate processes — identical mask hash,
scores, and polygon hash.

What the red team caught

Two Critical problems in the first pass, both fixed:

1. The headline metric was inflated by its own buffer. Bijection read 36/42 with a 14 px coastal
   buffer but 28/42 without, and 6 of the 10 anchors the buffer rescued were mainland, not
   offshore islands — segmentation holes being papered over. Buffer=0 is now the quoted gate;
   the buffered figure is a clearly-labelled secondary.
2. 44 of 64 junk polygons were water-coloured. The ocean rule convicted on isolation, which
   fails when SAM tiles open water into mutually-walled chunks. Now convicts on absolute background
   colour plus absence of interior land colour.

It also found the tests were pinning behaviour (n_bijective == 36) rather than constraining it —
they would have passed at 42 polygons or 103. Replaced with floors and property assertions.

Current honest numbers

map	territories	found	buffer=0 bijection
1 World Classic	42	85	33/42 (was 28)
79 Arctic Circle	34	139	—
100 Empires of the Med HD	150	283	—
34 Jungle of Despair	60	497	—

The 9 remaining World Classic failures are named individually in report.json. Five are
"label point in no polygon" for Britain, Indonesia, Western Australia, Alaska and Mongolia — the
genuine offshore-label case the buffer exists for. Three are hard pairs: Greenland/Iceland
(confirmed: D12 places Iceland's anchor physically on Greenland's landmass — unfixable from
artwork), South Africa/Madagascar, Irkutsk/Mongolia.

Still open

Over-production is the unsolved problem. 85 polygons for 42 territories; 497 for 60 on Jungle of
Despair, whose hatching fragments into foliage masks. The excess is title lettering, decorative
islands, coastal halo, and residual water.

Coastal halo was trimmed because it would otherwise manufacture false strait adjacencies in
Stage 3 — territories separated by narrow water appearing to touch.

Stage 2 (OCR/names) can proceed — real territories are outlined and contours are faithful.
Stage 3 (adjacency) cannot until over-production is solved: junk polygons would silently
produce false edges, and adjacency feeds every downstream metric.

Rejected on evidence: truncating to the catalog's num_territories. Keeping top-42-by-area
recovers only 32 of 39 real polygons, so it would mask real errors. num_territories is a warning
signal in the report, not a filter.