Skip to content

merge develop into ground-plan - #3481

Merged
timohueser merged 372 commits into
ground-planfrom
claude/gp-merge-develop-2
Nov 15, 2026
Merged

timohueser merged 372 commits into
ground-planfrom
claude/gp-merge-develop-2

Conversation

@timohueser

@timohueser timohueser commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner

Status (2026-11-14 evening)

Proofs on 1b9bb6f:

  • Plan path against ground-plan 46793ba.
    • 256 of 256 fixture digests are equal (records, faces, far, rim, mesh; storeys off as in the bridge).
    • The bridge WASM built from both trees is byte-equal (2,085,384 B).
    • 8 plan views: 7 are equal. underpass-eye differs by 579 px: the known two-state flip at the edges of one house, as before.
  • Default path against the last head 33897af: 12 views are equal to the pixel. ground plan GP12.3b part B: the viewer draws the decks, the device links the storeys #3482 changes only the plan path. The views: i01-wall-cap, i07-road-up-wall, main-road-15m, i09-shading, lake-shore, b315-bridge-junction, v17-sharp-corners, b45-bridge-junction, d-bridge-eye, under-bridge, junction0-far-700, stream-lake-far.
  • develop's stills on the new build, on the GPU: still.test.js 2 of 2 and home-views.test.js 6 of 6 (with @gpu) pass.
  • Checks:
    • obc test -p obc-trailview-features (122), -p obc-data-steps (80), -p obc-ground-plan (135).
    • clippy on these crates and cargo fmt --all --check.
    • obc suites check, obc prose --check, build_docs.py --check-links.
    • obc cairn map changes only the line, violation and problem counts, which check ignores. I did not commit them.
  • WASM guard in CI (run 43618609569):
  • CI: run 43618609569 on 1b9bb6f passes: 26 jobs pass, 2 are skipped. The wall time is 19.4 min (ios-app 18.6 min, test 11.9 min).
  • develop moved again after the re-merge, to ce7d966 (test: two cases of one test get their own names, so the test catalogue imports again #3485: one test name in roads/standing.test.ts). It merges without a conflict. I did not merge it.

The sections below describe the first merge (33897af). Their numbers hold for the code. The plan path numbers were measured against a8f1293.

Why

The owner wants ground-plan to merge into develop soon. This pull request is the preparation: it merges develop into ground-plan, so that the later merge into develop has no conflicts, and it shows what the default path (no ground=plan) gets. It does not change develop.

Commits merged

One merge commit, no rebase. Before the merge, develop had 361 commits that ground-plan did not have, and ground-plan had 655 commits that develop did not have. The merge base is 9ec892f.

Conflicted files

The merge gives conflicts in the 9 files of the audit (epic-audit/AUDIT.md section 4.2). I read the pull requests of both sides for each file.

File ground-plan develop Resolution
builder/web/src/lib/trailview/engine.ts The boot of #3225: start(canvas, gpu | Promise<Gpu>, ..., plan, finest, pool), engine.request() before the attach, the frames before the device (ask()), the canvas size in a method size(), eyeDrawn, the reach margin (#3260) The stills of #3400, #3404, #3450: start(..., scale, size?), a field size, no Controls and no ResizeObserver for a still, capture(), grab, copy(), the route ribbon lift LIFT_M, placed() Both. start(canvas, gpu, config, view, scale, { plan, finest, pool, size }): the options are one object, because two optional lists of positions can not both follow scale. The still's size is the field output (the name size is the method of ground-plan), and size() gives output when it is set, so ask() and fit() use the still's size. In the frame loop, the grabbed frame resolves first, then eyeDrawn runs as on ground-plan
builder/web/src/trailview.ts main() loads the release, the device, the bridge and the plan workers side by side; the pipelines compile at the attach open() gives the device and the release to main() and to stills() (?still), and calls prepare() Both. open(plan) follows the boot of ground-plan and gives the device as a promise. main() passes the promise to start. stills() waits for the device and calls prepare(), as on develop: no engine runs on a still page before the first still, so no worker waits for that compile
host/obc-data-steps/src/planner.rs trailview-features revision 18 revision 12 (#3274: the terrain reader of the deepest map pack) Revision 19 in 33897af; 20 after #3482 (see Status). Both sides used 9 to 12 for other contents, so the merged layer takes the next number above both sides at the time it lands. The merged bake code is that of ground-plan (it holds the same reader fix), and a bake of the fixture gives the committed tiles byte for byte (see Checks)
host/obc-trailview-features/src/terrain.rs The reader fix of #3274 with z == 13 The same fix with FINE_ZOOM develop's text. The tiles are at most zoom 13, so the two are the same rule
.github/workflows/ci.yml Downloads obc-ground-plan-bridge-pkg in web, web-browser and desktop-frontend Removes desktop-frontend and desktop; downloads obc-flat-device-pkg in web-browser; a shorter comment in web develop's jobs and comments, plus the plan bridge download in web and web-browser
tools/justfile The desktop recipe prints the per-checkout target Removes the desktop recipe Removed, as on develop
builder/web/vite.config.ts The plugins stampVersion(), bundleVersions() The VITE_FAKE_DEVICE define and a shorter licence comment Both. I also corrected the indent of develop's comment line
builder/web/src/lib/trailview/README.md The address rows ground and plan The address row still All three rows
docs/design/map.json Generated Generated Written again by obc cairn map. docs/design/README.md came out the same as on develop

Semantic conflict

svelte-check found one conflict that git did not show. ground-plan made View.plan a required field (the plan=faces debug draw). develop's coverView() (still-view.ts) and the fake renderer in covers/job.test.ts make views without it. Both now set plan: 'off', as defaultView() does. This is the one test file that I changed: the type of its fake view must match View.

I read the other files that both sides changed and git merged without a conflict: terrain/terrain.ts (develop: one comment line), harness.js (develop removes the FLAT ground, which no ground-plan code uses; ground-plan exports GPU), playwright.trailview.config.js (develop's serve.py, ground-plan's OBC_TRAILVIEW_GPU), tools/testing/suites.toml, tools/test_plan.py and its test, Cargo.toml, Cargo.lock (against ground-plan, only develop's own crates change; no third-party version changes), docs/design/trailview.md, docs/testing.md, AGENTS.md, specs/obc-data.md, host/obc-data-steps/src/functions.rs, host/obc-data/README.md, .gitignore, firmware/tools/dependency_rules.json and host/obc-trailview-features/src/roads/rows.rs (develop's clippy loop; the same as on ground-plan). Each holds the change of both sides, and no side uses a name that the other removed.

Notes:

  • With this merge develop loses the weekly TrailView job of test-weekly.yml (tests: cut the slow TrailView tests and route the census by its inputs #3216). The @gpu tests run in the manual suite web.trailview-gpu on a GPU. On SwiftShader, develop's still.test.js and home-views.test.js skip themselves anyway.
  • No ground-plan file names the desktop app that develop removed (grep desktop in the suites, the justfile and the CI files finds only sim/desktop).

Plan path (ground=plan): no change against ground-plan

The plan path of the merge is the plan path of ground-plan:

  • Code. host/obc-ground-plan, builder/ground-plan-wasm, data/trailview and the browser fixture are byte-equal to ground-plan. Against ground-plan, the merge changes only Cargo.lock by develop's own crates (no third-party version moves). The plan bridge that this PC builds from the merge is byte-equal to the one it builds from ground-plan (2,002,512 B). In CI the bridge is 772,451 B gzip and 2,001,570 B raw, as in the run of ground plan GP12.3b part A: the levels in the bake, the storey shell without its own machinery (WASM cuts) #3470.
  • Digests. The 256 fixture cells, by the bridge's own path (build_request, then mesh_cell with the rims of the cells about it), storeys off as in the bridge: records, face words, far block, rim and mesh digests of ground-plan a8f1293 and of the merge are equal in 256 of 256 cells.
  • Fixture. A bake of the fixture box with the merged trailview-features (the inputs of baseline/feature-bake, 4 threads) gives the committed tiles byte for byte (make-fixture.py features: no file changes). So revision 19 needs no new fixture bytes; the revision only keys the next obc data bake.
  • Views. 8 reference views with ground=plan (v17-sharp-corners, b45-bridge-junction, d-bridge-eye, underpass-eye, stream-culvert-main-road, lake-shore, junction0-eye, i07-hairpin-eye), GPU, 1920 × 960, the same fixture data: 7 views are equal to the pixel. underpass-eye takes one of two states from run to run on both builds (3 runs each: the states differ by 579 pixels at the edges of one house); each state of one build is equal to a state of the other.

Default path (no ground=plan): what develop's users get

Method: every view of road-views.json (120), road-views-top.json (50), road-views-far.json (18) and views.json (3), shot on the GPU (GTX 1660 SUPER, under the lock, 1920 × 960, one page at a time) from fixture builds of develop 636f9ce and of the merge 33897af. Both builds read the same data: the fixture terrain and the merge's fixture feature tiles, no ground layers. A pixel counts when one channel differs by more than 8 of 255. To name the causes, I shot the default path of ground-plan's first-parent history (79 merges) with the merge's harness and the same data: 24 sample views at every 10th merge and then at every merge of each interval that changed, all 191 views at the merge after each cause, and the changed views at the merge before it. The shots are on timosPC: /media/timo/OSM1/tmp/ground-plan/develop-merge-2/ (sheets/INDEX.md; the per-view numbers in sheets/default/attribution2.tsv).

Result:

  • ground-plan a8f1293 | merge: 191 of 191 views equal. Nothing on the default path comes from the merge itself.
  • develop | merge: 188 of 191 views differ, 3 are equal. Every part of 100 pixels or more is named below. The merge's default path gave the same pixels in two runs (36 views); develop and the early ground-plan builds flip between two states in some views from run to run (for example v11-two-bad-junctions 1,533 px, b45-bridge-junction-far-700 1,381 px, conn-path-off-track-eye 760 px): such parts are named "flip".
Pull request Views (part ≥ 100 px) What changes on the default path In the audit's list (section 4.3)
#3188 walls + D1 177 data/trailview/roads.json: the cut and fill slopes of every class (roads 0.85 → 2:3, tracks 0.85 → 1:1, paths 2.14 and 0.70 → 1:1), from its commit 37664cd. The old corridors read these slopes, so the banks beside every road and path are gentler and wider, wall tops are lower, and the ground under an eye moves, so top views move a little. Proof: ground-plan just before #3188 with only this file changed gives the shots after #3188 to the pixel in 10 of 10 views No. The audit lists #3188 as "0 pixels on the default path"; its body compared against 35c8f49, which had the new slopes already
#3224 look thresholds 156 Light bands and cut rock on slopes no sharper than their data: softer bands on grass slopes Yes
#3219 cut and fill look 149 The cut and fill bands beside roads draw with the terrain's own look (no mottled cut texture) Yes
#3385 window distance 117 Windows on houses stay to twice the distance (far views, top views) Yes
#3226 far lake lift 18 Water lift from the point's LOD: small edges, mostly in far views (104 to 1,185 px) Yes
#3164 GP9 live 15 Depth ties on corridor edges (110 to 1,024 px; its body: "0.8 to 1.8 % of pixels", corridor edges) No (the audit names #3179 and #3180 for the old path)
#3215 water A 10 FOAM_MOST 0.35 on the stream ribbons: less white foam on streams (411 to 16,281 px, for example b315-bridge-junction 15,281 px) Yes
#3160 GP7-C 1 tears-zigzag: 2,191 px at a far road in the top left corner No

Worse in places (judgement for the owner, from the sheets):

#3179 and #3180, which the audit lists for the old path, change no view by 100 pixels or more here (the merges #3167 to #3204 change only flipping views).

Not visible in these shots: the signs on buildings (#3315; the fixture tiles have no use tag, so develop draws none either), and the load order and caches (#3225, #3263, #3322, #3323, #3268, #3220: an idle shot shows the final image only). The data change of revision 19 (road and water heights of the new solver) is not in these shots by design; it comes with the next bake.

Per view: pixels over 8 and the named parts
View px > 8 Named (px of each part)
i01-wall-cap 260056 #3188 280497; #3219 136210
i02-dark-lines 42572 #3164 723; #3188 17764; #3219 17523; #3224 6742; #3385 4793
i03-guardrail 51578 #3188 34095; #3219 22891; #3224 1539; #3385 8993
i04-markings 4420 #3164 616; #3188 711; #3224 2984
i05-dark-shadow 13860 #3164 128; #3188 4243; #3219 1336; #3224 3306; #3385 5624
i06-trail-width 96223 #3188 93781; #3219 2232; #3224 4011; #3226 125
i07-road-up-wall 507850 #3188 35069; #3219 472940; #3224 60576
i08-river 82240 #3188 43808; #3219 58410; #3385 6856
i09-shading 78877 #3188 11772; #3219 18423; #3224 203; #3385 53590; #3226 124
mouth-stream 10854 #3188 3163; #3219 5537
mouth-stream-far 20987 #3188 8714; #3219 2979; #3224 382; #3385 9085
mouth-track 80588 #3188 38638; #3219 41483; #3224 5452; #3385 16292
i10-intersection 20564 #3188 4442; #3219 5472; #3224 3987; #3385 6123
main-road-eye 251560 #3188 181916; #3219 199933; #3224 495; #3385 370
main-road-15m 844756 #3188 258555; #3219 800107; #3224 3898
side-eye 188481 #3188 186833; #3219 106343; #3224 502; #3385 209
side-15m 85482 flip (about 175 px) 121; #3188 74928; #3219 64674; #3224 107; flip (about 175 px) 173; #3385 644; #3226 175
hp-eye 228901 #3188 156317; #3219 165628; #3224 807
hp-15m 112060 #3188 79462; #3219 67768; #3224 2242; #3385 1695
d-hairpin-eye 142515 #3188 84919; #3219 101376; #3224 9651
d-hairpin-15m 99575 #3188 61268; #3219 57072; #3224 8872
secondary-eye 25480 #3188 1425; #3219 5164; #3224 674; #3385 18426
mountain-eye 168798 #3188 148226; #3219 73945; #3224 10871
valley-eye 30425 #3188 9224; #3219 20999; #3224 5542; #3226 181
residential-curve-eye 13397 #3188 6508; #3219 1302; #3224 1359; #3385 1951
residential-15m 71440 #3188 44990; #3219 13803; #3224 6821; #3385 810
residential-night 61365 #3188 58227; #3385 3097
junction0-eye 92906 #3188 66841; #3219 22462; #3224 808; #3385 3257
t-tertiary-eye 23688 #3188 1199; #3219 6218; #3385 16458
x-tertiary-eye 15204 #3188 1118; #3224 165; #3385 13933
t-mouth-close 23913 #3188 949; #3219 4108; #3224 898; #3385 18086
crossing-close 39700 #3188 6182; #3219 6472; #3224 356; #3385 27042
sett-eye 7839 #3188 858; #3385 6957
steps-town-eye 8895 #3188 1570; #3219 229; #3385 7210
d-bridge-eye 20608 #3188 5654; #3219 145; #3224 1597; #3385 13684
d-bridge-15m 19860 #3188 3271; #3219 321; #3224 685; #3385 15080
underpass-eye 107173 #3188 68696; #3219 77504; #3224 2560; #3385 4876; #3226 148
under-bridge 236496 #3188 219261; #3219 160949; #3224 1124; #3385 727
bridge-eye 3375 #3188 453; #3219 120; #3224 2170
gravel-eye 29951 #3188 2614; #3219 13264; #3224 17109; #3385 142
dirt-eye 42040 #3188 5726; #3219 29888; #3224 19930
grass-eye 72995 #3188 56559; #3219 62568; #3224 431
hairpin-eye 37231 #3188 21934; #3219 27304; #3224 216
trail-eye 0 equal
alpine-trail-eye 27872 #3188 26272; #3224 532
footway-eye 31753 flip (1476 px, early builds flip run to run) 1476; #3188 26634; #3219 854; #3224 228; #3385 2554
cycleway-eye 47637 #3188 37292; #3219 37018; #3224 249
b-steps-eye 4114 #3188 1292; #3219 1499; #3224 527; #3385 1012
rail-narrow-eye 176474 #3188 70916; #3219 104904; #3385 2951
rail-close 29464 #3188 7393; #3219 186; #3224 1312; #3385 21058
funicular-eye 2253 #3188 1788; #3219 445; #3224 116
lake-shore 65159 #3188 24798; #3219 66434; #3224 678
river-area-track 110283 #3188 75230; #3219 100428
zebra-tagged 1511 #3188 1489
path-on-cut 36462 #3188 32836; #3224 1163
j-tt-a 7266 #3188 7267
streams-under 14208 #3188 1179; #3219 9681; #3224 4223
streams-under-far 10098 #3188 1199; #3219 4572; #3224 402; #3385 4261
stream-culvert-main-road 7440 #3164 186; #3188 1312; #3224 6082
stream-culvert-track 10159 #3188 2077; #3219 4207; #3224 6107
stream-plain-track 8930 #3188 3437; #3219 1174; #3224 5951
stream-under-bridge 86090 #3188 44954; #3219 60652; #3224 8632
tears-zigzag 2910 #3160 2191; #3188 3903; #3224 354; #3385 595
tears-zigzag-90m 113295 flip (about 325 px at every step) 325; #3188 36161; flip (about 325 px at every step) 331; #3219 89439; #3224 3444; #3385 4199; #3226 474
forecourt-top 472 #3224 159; #3385 265
stream-bridge-far 921 #3188 367; #3219 133; #3224 121; #3226 238
stream-lake-far 2706 #3215 2426; #3226 147
stream-parallel-road 10378 #3188 7291; #3224 2194; #3385 755
conn-own-a 79556 #3188 39884; #3219 38749; #3224 5092; #3385 20311
conn-own-b 69770 #3188 83660; #3219 55955; #3385 4411
conn-own-c 89282 #3188 98312; #3219 41070; #3224 2355; #3385 1857
conn-path-off-road-above 259041 #3164 562; #3188 258938; #3219 31188; #3224 3749
conn-path-off-road-eye 314622 #3188 273271; #3219 61851; #3385 508
conn-track-off-road-above 107738 #3164 192; #3188 34385; #3219 88305; #3224 8175; #3385 432
conn-track-off-road-eye 38520 #3164 1024; #3188 8425; #3219 24701; #3224 2654; #3385 5172
conn-path-off-walk-above 103641 #3188 118097; #3219 102059; #3224 1071
conn-path-off-walk-eye 33144 #3188 27009; #3219 38028; #3224 188
conn-road-into-path-above 25876 #3164 311; #3188 19433; #3219 6728; #3224 1674; #3385 704
conn-road-into-path-eye 122979 #3164 345; #3188 11395; #3219 104849; #3224 615; #3385 6586
conn-walk-into-path-above 112575 #3188 119155; #3219 64819; #3224 676
conn-walk-into-path-eye 121547 #3188 140670; #3219 103787
conn-road-into-track-above 39131 #3188 27439; #3219 5734; #3224 7575
conn-road-into-track-eye 153798 #3188 121679; #3219 3678; #3224 22853; #3385 5600
conn-walk-into-track-above 43613 #3188 19278; #3219 24759; #3224 1261; #3385 426
conn-walk-into-track-eye 13864 #3188 6071; #3219 6280
conn-path-off-track-above 175902 #3188 61767; #3219 143642; #3224 7318
conn-path-off-track-eye 45280 #3188 13352; flip (761 px; develop flips 760) 761; #3219 24155; #3224 1746; flip (761 px; develop flips 760) 761; #3385 5915; #3226 114
path-edge-eye 62700 #3188 57448; #3219 1281; #3224 4595; #3385 4305
path-edge-handover 12019 #3188 4957; #3219 7978; #3224 524
stream-mouth-eye 6048 #3188 906; #3219 3140; #3224 1237; #3385 835
stream-mouth-eye-north 22691 #3188 11926; #3219 106; #3224 10016; #3385 7740
stream-main-road-eye 38232 #3164 146; #3188 14334; #3219 320; #3224 2557; #3385 20652
conn-footway-hump-above 83610 #3188 84573; #3219 46387; #3385 1012
conn-footway-hump-eye 64033 #3188 33035; #3219 51024; #3385 5529
conn-steep-ramp-above 297229 #3188 211235; #3219 282929; #3224 2540
conn-steep-ramp-eye 252233 #3188 223036; #3219 327552
conn-parallel-way-above 9339 #3188 6765; #3219 1263; #3224 1964; #3385 443; #3215 411
conn-parallel-way-eye 14057 flip (127 px) 132; #3188 4228; #3219 1388; #3385 8247
i07-hairpin-eye 247779 #3188 150772; #3219 78676; #3224 16394; #3385 26494
worst3 184289 #3188 86793; #3219 157493; #3224 1280
flat-bend-top 114846 #3188 85579; #3219 60467; #3224 8003; #3385 311
v01-stairs 5927 #3215 5847
v02-stream-path-shrubs 16481 #3188 25040; #3215 16281
v03-good-path 588 #3385 531
v04-wall-road-junction 163342 #3188 132957; #3219 112825; #3224 5733; #3385 553
v05-slope-artifact 11835 #3188 8569; #3219 2206; #3224 117; #3385 400
v06-road-path-broken 46895 #3188 16568; #3219 34523; #3224 7839
v07-good-junction 43070 #3188 12277; #3219 36943; #3224 6088
v08-rough-connection 23762 #3188 1794; #3219 17084; #3224 7455; #3385 572
v09-sidewalk-patches 4338 #3188 784; #3219 1659; #3224 865
v10-signpost-bench 586 #3188 511
v11-two-bad-junctions 18736 #3164; develop flips 1533 px run to run 1836; #3188 5119; #3219 7533; #3224 4807; #3385 921
v12-good-town-junctions 7793 #3164 549; #3188 6369; #3224 939
v13-culverts 12758 #3188 184; #3224 11627; #3385 471
v14-paths-not-joined 18820 #3188 565; #3224 2521; #3385 15682
v15-stuff-on-road 24267 #3188 914; #3224 1453; #3385 21846
v16-road-rail 13096 #3188 1586; #3224 2814; #3385 8696
v17-sharp-corners 159604 #3188 80333; #3219 116110; #3224 14948; #3385 370
b45-bridge-junction 5885 #3188 2445; #3219 511; #3224 2240
b315-bridge-junction 187884 #3188 179768; #3219 16524; #3224 3293; #3215 15281
v01-stairs-top 12341 #3188 1078; #3215 12548
v02-stream-path-shrubs-top 3730 #3188 286; #3215 3923
v03-good-path-top 1801 #3188 1681
v04-wall-road-junction-top 64747 #3188 61634; #3219 43969; #3224 1921
v05-slope-artifact-top 21755 #3188 17984; #3219 9804; #3224 344
v06-road-path-broken-top 58982 #3188 20173; #3219 38214; #3224 5441; #3385 3498
v07-good-junction-top 20899 #3188 9707; #3219 17502; #3224 2047
v08-rough-connection-top 16995 #3188 4328; #3219 10762; #3224 2692
v09-sidewalk-patches-top 4548 #3188 159; #3219 1700; #3224 991; #3385 605
v10-signpost-bench-top 2572 #3188 171; #3224 1035; #3385 1492
v11-two-bad-junctions-top 8337 #3188 1247; #3219 3629; #3224 1047; #3385 1824
v12-good-town-junctions-top 31293 #3188 1703; #3219 882; #3224 2641; #3385 25970
v13-culverts-top 2173 #3224 2020
v14-paths-not-joined-top 3333 #3188 612; #3224 1380; #3385 1351
v15-stuff-on-road-top 41447 #3188 1874; #3224 3984; #3385 35595
v16-road-rail-top 39035 #3188 507; #3224 3280; #3385 35170
v17-sharp-corners-top 92443 #3188 53746; #3219 45658; #3224 5459; #3385 15516
i07-hairpin-top 86538 #3188 52748; #3219 46277; #3224 4692; #3385 11441
d-hairpin-top 122762 #3188 46417; #3219 102727; #3224 4102
hp-top 122015 #3188 80466; #3219 80021; #3224 4385
t-tertiary-top 6558 #3188 576; #3219 563; #3224 2072; #3385 3553
x-tertiary-top 3239 #3224 638; #3385 2435
junction0-top 83136 #3188 50919; #3219 53867; #3224 1594; #3385 3164
j-tt-a-top 3690 #3188 3595; #3224 146
b45-bridge-junction-top 7141 #3188 3330; #3219 1916; #3224 439; #3385 2016; #3215 824
b315-bridge-junction-top 7634 #3188 4803; #3219 1985; #3224 406; #3385 1120; #3215 845
d-bridge-top 81678 #3188 1271; #3219 1154; #3224 3786; #3385 75073
zebra-tagged-top 5626 #3224 334; #3385 5196
crossing-close-top 36665 #3188 2508; #3219 2004; #3224 1132; #3385 31806
stream-culvert-main-road-top 2150 #3224 2020
stream-culvert-track-top 8516 #3188 1486; #3219 4603; #3224 4314
mouth-stream-top 10844 #3188 3786; #3219 1899; #3224 767; #3385 4594
stream-parallel-road-top 11164 #3188 1882; #3224 2213; #3385 6993
lake-shore-top 58204 #3188 54150; #3219 51611; #3224 1106
rail-close-top 19989 #3188 842; #3224 919; #3385 18137
steps-town-top 18752 #3188 5960; #3219 1324; #3224 570; #3385 11341
conn-own-a-top 32378 #3188 17191; #3219 21587; #3224 1756; #3385 1977
conn-own-b-top 23136 #3188 23243; #3219 7397; #3385 5749
conn-own-c-top 39456 #3188 37241; #3219 7051; #3224 522; #3385 6631
conn-path-off-road-top 44098 #3164 110; #3188 34109; #3219 11145; #3224 2281; #3385 11807
conn-track-off-road-top 185428 #3164 137; #3188 87207; #3219 120111; #3224 11002; #3385 2919
conn-path-off-walk-top 38980 #3188 37716; #3219 20592; #3224 2546; #3385 7627
conn-road-into-path-top 37449 #3164 155; #3188 20395; #3219 19825; #3224 1207; #3385 1619
conn-walk-into-path-top 42281 #3188 37759; #3219 7396; #3224 634; #3385 9632
conn-road-into-track-top 25393 #3188 13411; #3219 3844; #3224 7907; #3385 1805
conn-walk-into-track-top 24426 #3188 9126; #3219 10180; #3224 1631; #3385 4116
conn-path-off-track-top 190944 #3188 66115; #3219 153193; #3224 7290
conn-footway-hump-top 26452 #3188 15218; #3219 6027; #3224 543; #3385 9287
conn-steep-ramp-top 61468 #3188 50852; #3219 50314; #3224 1783; #3385 161
conn-parallel-way-top 8902 #3188 3783; #3219 655; #3224 3212; #3385 1581; #3215 438
junction0-far-300 28044 #3188 6774; #3219 659; #3224 338; #3385 21299
junction0-far-700 73486 flip (258 px) 272; #3188 8131; #3219 1209; #3224 123; #3385 64382; #3226 104
junction0-far-1400 15689 flip (1521 px) 1521; #3188 7839; #3219 6177; #3224 161; #3385 641; #3226 301
j-tt-a-far-300 25473 #3188 13430; #3219 19216; #3224 645; #3385 142
j-tt-a-far-700 11238 #3188 4206; #3219 8344; #3224 275
j-tt-a-far-1400 0 equal
t-tertiary-far-300 94326 #3188 2213; #3219 5360; #3224 435; #3385 86586; #3226 255
t-tertiary-far-700 6413 #3188 2825; #3219 2000; #3224 406; #3385 965; #3226 306
t-tertiary-far-1400 0 equal
x-tertiary-far-300 60626 #3188 3890; #3219 1187; #3224 163; #3385 55317; #3226 237
x-tertiary-far-700 28700 flip (482 px; develop flips 481) 484; #3188 1694; #3219 1715; #3385 24342; #3226 284
x-tertiary-far-1400 2577 flip (about 290 px) 304; #3188 1029; #3219 549; #3385 520
i07-hairpin-far-300 173881 #3188 9133; #3219 2591; #3224 150; #3385 162997
i07-hairpin-far-700 33675 #3188 5592; #3219 1448; #3224 254; #3385 26365
i07-hairpin-far-1400 4396 #3188 2037; #3219 983; #3224 205; #3226 1185
b45-bridge-junction-far-300 30623 flip (591 px; develop flips 591) 591; #3188 3212; #3219 4209; #3224 287; #3385 23685
b45-bridge-junction-far-700 55526 flip (1381 px; develop flips 1381) 1381; #3188 1385; #3219 1453; #3224 205; #3385 51119
b45-bridge-junction-far-1400 6327 flip (1475 px) 1475; #3188 3023; #3219 855; #3224 752; #3226 284
walk-engelberg-to-titlis 11410 #3385 11354
fly-12km-over-engelberg 374 #3224 161; #3226 212
fly-60km-above-the-fixture 78 parts under 100 px only: #3224 43, #3457/#3449 or #3226 or #3464/#3470 35

Measurements

Main bundle of the default path (fixture builds of the TrailView page, vite build --mode trailview, raw / gzip -9):

Chunk develop 636f9ce merge 33897af Growth
TrailView page entry trailview-*.js 860,795 / 280,191 B 1,142,498 / 363,657 B +281,703 / +83,466 B (+32.7 % / +29.8 %)
All JS of the TrailView page (entry and workers) 1,460,350 / 509,879 B 1,810,137 / 603,600 B +349,787 / +93,721 B
Planner entry planner-*.js 59,205 / 22,907 B 59,268 / 22,923 B +63 / +16 B
Planner's lazy TrailView chunk engine-*.js 339,448 / 121,647 B 434,749 / 153,819 B +95,301 / +32,172 B (+28 % / +26 %)

ground-plan a8f1293 gives 1,131,858 / 358,687 B for the TrailView entry: the merge adds develop's stills (+10,640 B raw). The plan bridge (2.0 MB raw, 767 KB gzip) is a separate asset in both builds; the default path does not fetch it.

WASM guards in CI (run 43527050344): plan bridge 772,451 B of 773,120 B gzip (669 B left) and 2,001,570 B of 2,001,920 B raw (350 B left), the same as ground-plan's run for #3470; builder core 496,290 B of 555,008 B gzip and 1,288,539 B of 1,359,872 B raw, the same as develop's push run 43521754194.

CI wall time (first job start to last job end): this pull request 18.6 min (25 jobs; ios-app 17.7 min, test 15.2 min, of which cargo nextest fast 9.3 min and fixtures 4.2 min). develop pull requests that change Cargo.lock and so run the whole Rust workspace: #3447 14.0 min (test 7.2 min, fast 2.4 min), #3402 18.5 min (test 8.8 min, fast 2.9 min). So the test job takes about 6.5 to 8 minutes more; its fast tier now also builds and runs obc-ground-plan with its fixture tests. develop push runs that select less take about 7 min (636f9ce: 7.5 min).

What the code says about older tiles

I did not download or open a production release. What the code says:

  • The viewer opens a feature archive by its metadata, not by the layer revision: featureTiles (stream/terrain.ts) reads an archive with features_format 4 and shows no features, with a warning, for another format. Both sides keep format 4 (stream/features.ts FORMAT). The revision (12, 18 or 19) is a key of obc data only: no tile and no metadata holds it. So by the code, the merged viewer reads the feature tiles of a release baked at revision 12 on the default path. The tags that ground-plan added (levels, storeys, crossing states, shallow tunnels) are read only on the plan path.
  • develop's releases have no trailview-stamps layer (develop's catalog does not name it). Without it, or with another stamp_version, bakedNodes gives nothing and the stamp worker makes every node, as on develop today.
  • The ground, colour and snow layers keep their formats (stream/ground.ts FORMATS: 3, 2, 1) on both sides.
  • Which revision today's production release holds: not checked.

Checks

Run on this PC, own target checkouts/429fbe719546aee5, capped, all pass:

  • obc suites prepare (the three WASM bridges built from this checkout).
  • obc test -p obc-trailview-features (122), -p obc-data-steps (80), -p obc-data (262, 1 skipped), -p obc-ground-plan (135, with the_tiles_carry_the_levels_of_each_cell).
  • cargo clippy -p obc-data-steps -p obc-trailview-features -p obc-data -p obc-ground-plan --all-targets -- -D warnings; cargo fmt --all --check.
  • builder/web: npm run check (0 errors, 0 warnings), npm run check:wgsl, vitest on trailview/sight.test.ts, still-view.test.ts, terrain/terrain.test.ts, terrain/model.test.ts, plan/plan.test.ts, view.test.ts, terrain/stamp-sources.test.ts, src/lib/covers and src/lib/platform/bundle.test.ts (11 files, 79 tests).
  • On the GPU against the merged build: still.test.js (2 tests) and home-views.test.js (6 tests, with the @gpu render): develop's stills work after the merge. The plan boot works: the ground=plan shots above build their records (35 to 56 cells per view) with no page error.
  • obc suites check, obc prose --check, python3 docs/build_docs.py --check-links, obc cairn map.
  • The fixture bake and the 256 digests above.

CI: run 43527050344 on 33897af, all jobs pass.

Left out locally: the Python suites python.repository-tools and python.trailview, the rest of obc test affected (the whole Rust workspace, because Cargo.lock changed), the SwiftShader browser suites, firmware and iOS. CI ran them.

Changed tests: one. builder/web/src/lib/covers/job.test.ts: its fake renderer makes a View, which now needs plan: 'off' (see Semantic conflict).

What is left

  • The owner's go for the merge into develop. He accepted the default-view changes (the slope change of #3188, the banks without the cut texture of #3219, and the other changes of the table above) on 2026-11-14.
  • An adversarial review of the three merge commits.

Public docs

No page of docs/content/ changes in the merge commit beyond what develop already changed. Against ground-plan, the public docs that change all come from develop. The merge edits two generated or reference files only: the TrailView README rows (one row from each side) and docs/design/map.json (written by obc cairn map).

Mode: change #3096
Requirements: none

🤖 Generated with Claude Code

timohueser and others added 30 commits November 12, 2026 16:32
terrain: wait for Kartverket's overuse answer and pace Norway's requests
…f a line

still() starts an engine of its own on a canvas that the page does not
show, at a size in pixels, and ends it after the still. The engine draws
the jittered frames of the view again and copies the canvas and the depth
of the last frame. For a line, it puts each point on the drawn terrain as
the route ribbon does and tests it against that depth. coverView() chooses
the view of a cover from at most four terrain tiles. With `still` in the
address, the page starts no engine and gives both functions to a script.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A duplicate keeps the summary of its source. The planner counts and removes
records that are not route records. The store reads indexedDB only when it
opens. The state checks use only own keys.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Delete builder/desktop and the web code that only the desktop host
reaches: the device page, the ride library page, the desktop page, the
gating layer and the desktop platform. The hosted map builder keeps the
device step; a browser with no WebUSB now shows "This browser cannot use
the USB cable. Use Chrome or Edge."

obc-usb stays for `obc format-card`. Its `open` returns only the link,
and the Linux udev rule moves to builder/usb/linux with a README.

CI loses the desktop-frontend and desktop jobs, and the tools, suites,
licences and data readers lose their desktop entries.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The new crate host/obc-fit reads the FIT container in both byte orders,
with compressed timestamps, and skips developer fields and unknown
messages. A cut file or a wrong CRC gives the records before the damage.
It writes an activity file for a ride and a course file for a route.

The bridge exports obc_fit_read, obc_fit_write_ride and
obc_fit_write_course, and lib/core/fit.ts wraps them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Library store, summaries and the backup file
…-desktop-app

# Conflicts:
#	docs/design/map.json
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…draw

An engine with a size makes no controls and no resize observer, and only it copies the canvas and the depth. A capture rejects when no frame can draw. A route of no extent gets no view. The GPU test skips itself on a software renderer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ecords

A course point of a turn type is no waypoint. The serial number of a
course comes from a hash of its name, positions and time, so two courses
have two file_ids. A compressed timestamp past the end of u32 is
dropped. A course keeps a record with a position and no time. The
fixture test waits for the owner's files and moves to a later change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Delete the three-way confirm, the TypeScript trip codec, and the OBCR
waypoint reader with its wasm export. `vendorIn` and `siteNav` are now
required, so the "unavailable" error and the header without site links
go. The device end of the loopback link is a `RecordLink`. Comments no
longer name the desktop app.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
trailview: render one view to a still, with the place of each point of a line
The still gives its credits. The script home-views.mjs renders the views
of home-views.json at times by the sun, uploads a set under its digest,
and puts the manifest of a set at the live key.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…till

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The cover job renders the TrailView still of a route or a ride, one at a
time in the whole browser, and stores the image and the place of each
point of the summary line in the library. Cover.svelte draws the line on
the cover, with a mark and a selected part, or the line on a plain ground
when there is no cover. planner.ts makes the one library of the page,
gives it to the planner and starts the job.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…over

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
switch checks the bucket, not the cache of the public host. upload resolves
--out and takes only images of the production release. A list takes the
whole copied address. A sun up at midnight on the Zurich clock gets its own
refusal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
#	docs/design/map.json
Keep an ask that comes during a render or before the job starts, store
no cover for a deleted record, keep far points as floats, and make the
screenshot test fail on a lost line, mark or hidden strength.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rust 1.100 extends needless_range_loop and needless_bool. Seven places
on develop now fail clippy with -D warnings. Each change keeps the
behavior.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
timohueser and others added 12 commits November 13, 2026 20:32
…tings

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Ride page: one line large in a window, in the colours of its zones (#3433)
The name of the axis beside the ticks switches between the distance and
the time of day. A stop is a mark on the row of the ticks, and the foot
of the pointer says its length. With a part selected, the menu "More"
has "Only the selected part" for the three exports. With every line
hidden, the card shows its chips only.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The cover drew a selected part in amber, and the map drew it in the colour
of the line. Now both draw the part in the colour of the line and the rest
pale. On the cover the pale rest has no glow and no core: with them, the
light tint was brighter than the part on a dark slope.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… card as the part

The foot of the pointer on the time axis says the time at the place of
the pointer. The length of a stop rounds to minutes before it changes to
hours. The stops are found one time for a ride. A click on a climb in
the left card selects it as the part.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The rest of the line draws before the part, so the part is on top where
the line crosses itself or goes back on itself. A segment that overlaps
the part is in the part, so a part shorter than one segment shows. The
test reads the colour, the strength and the order of each stroke.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Ride page: the time axis, the stops and the export of a part (#3433)
…way back

A drag across the line of the window shows that stretch large. A pinch on
a trackpad changes the length of the view, and a sideways scroll moves it.
The ticks, the scale, the numbers and the zone times are of the view. The
button "Whole ride" stands beside the title while the view is a stretch;
Escape goes back to the whole ride first and closes the window second. The
window opens on a selected part.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Cover: a selected part keeps the colour of its line, as on the map
A sideways scroll on the line never reaches the browser, also in the whole
ride. The numbers of a stretch are of the samples from the first to the
last one in it, and of the two samples around a stretch with none. While
the view is that of the selected part, the numbers are of the samples of
the part. The animation stops with the window, the scale of the whole line
is computed once, and the ticks are fine only in a stretch.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Ride page: a stretch of one line large in its window, and an obvious way back
Merges origin/develop 636f9ce into ground-plan a8f1293. Conflicts:
- engine.ts, trailview.ts: develop's stills and the plan's boot both stay.
  `TrailView.start` takes its options in one object (plan, finest, pool,
  size); the still's output size goes through `size()`, which the frames
  before the device and `fit()` read.
- planner.rs: trailview-features takes revision 19, above develop's 12 and
  ground-plan's 18, which both used 9 to 12 for other contents.
- terrain.rs: develop's reader with FINE_ZOOM (the same rule on both sides).
- ci.yml: develop's jobs, without the desktop jobs, plus the ground plan
  bridge downloads.
- justfile: the desktop recipe goes, as on develop.
- vite.config.ts: develop's define, plus the stamp and bundle plugins.
- README: both rows. map.json: written again by `obc cairn map`.

Semantic conflict: ground-plan's `View` requires `plan`, so develop's
`coverView` and the fake renderer of `covers/job.test.ts` set `plan: 'off'`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@coderabbitai

coderabbitai Bot commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: 1be0cd9c-f9ed-4a21-922d-7d731a677251

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • Trigger review
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

timohueser and others added 5 commits November 14, 2026 00:46
The state of a backup has the field `nextMap`: the regions, the boxes and
the two map styles of the device map that the rider chose, and the size of
its download.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…s the map that waits

The page `#/device/area` shows the coverage map with the ground outside the
area pale and the routes of the library on it. The rider adds regions with a
click and boxes with a drag, sees the routes that leave the area with two
answers, opens the map styles in a window, and Done writes the library state
`nextMap`. The page works with no device.

The row Map of the Device page shows the map on the device, or "New area" or
"New colours" with the size of the download, and the link to the page. The
button "Send the map" is in the row and does nothing yet: no code builds a
map.

The coverage map has the props `tool`, `rail`, `pale` and `routes`, and it
loads its own style rules, so a page of the shell can show it.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- The Device page loads the regions of the map catalog one time, and the
  row Map keeps its text during a sync.
- Done compares the map with the stored `nextMap`, so a stored map with a
  region that the catalog no longer holds is written again.
- A route that leaves the area on two sides gets one box for each side.
- A route with no length outside the area gets no warning.
- A stored box must have a size.
- The dark rule of the pale ground repeated the light value; a comment
  named a rule that is not there; an empty branch is gone.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The button of the row Map starts a run of its own in the device sync state:
it downloads the cells of `nextMap`, assembles the map in the tab and puts
it on the device. A sync never builds a map, and the sync and the map send
wait for each other. After the put the run remembers the area of the map
object and the map waits no more, unless the rider chose another map during
the build.

While the run goes, the row shows one line with the step and one bar, and
"Stop" in the place of the button. Stop, a pulled cable and a card with no
room keep the old map on the card and the new map waiting, each with one
sentence. A browser that saves data asks first.

The working files of map assembly have the Web Lock `obc-map-work`, with a
signal that ends a wait, so the map builder page and this run do not delete
the files of each other.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- The build makes exactly the map that waits or nothing: a region, a box
  or a map style that is not there any more refuses the run, so the
  remembered area of the map on the device is never false. The comparison
  of two maps is now in `lib/device-sync/area.ts`.
- Stop is not available while the device stores the map.
- Stop and a pulled cable also end the wait for the modules and for the
  map catalog.
- A failed write of the library after the put is no failed send.
- The sentence of the last send goes when another map waits.
- The text for a send that did not end makes no claim about the download.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@timohueser
timohueser marked this pull request as ready for review November 14, 2026 06:46
timohueser and others added 3 commits November 14, 2026 09:31
…intly

The ground outside the area lost its contrast under the pale layer. It is
now grey with a light veil: the names, the roads and the lakes keep their
contrast, and the area stands out by its colour.

The lines of the map cells show faintly where a cell is at least 36 pixels
wide, in both modes: an area grows to whole cells, and the lines show why.
While the rider draws a box, a line goes around the cells that it takes.

Both are for a page of the shell (`pale`). The map builder page is as it
was.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… draws only what the renderer shows

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Device map: the page for the area and the colours, and "Send the map" (#3441)
@timohueser

Copy link
Copy Markdown
Owner Author

Owner decision, 2026-11-14, on the changes to the default view (the table of this pull request and the sheet develop-merge-2/sheets/kinds.png): "yeah those changes are fine! let's move on!"

The default-view changes are accepted. Still needed before ground-plan goes into develop: a review of this merge commit, one more merge of origin/ground-plan (46793ba; the features layer then takes revision 20), and the owner's explicit go for the merge and the push.


Generated by Claude Code

timohueser and others added 2 commits November 14, 2026 17:51
Merges origin/ground-plan 46793ba (#3482, deck step part B) into the merge
of develop. Both sides had trailview-features revision 19 for other
contents (this branch: develop's 12 and ground-plan's 18 merged; #3482: the
merged A lines), so the layer takes the next number above both: 20.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Merges origin/develop 19dc9d0 (#3483, the device's area page), which moved
after the first merge. No conflict; it touches no TrailView file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@timohueser

Copy link
Copy Markdown
Owner Author

Review of PR #3481 (develop into ground-plan), head 1b9bb6f

Verdict: MERGE (into ground-plan). No blocking finding. Items N1 to N4 must be settled before the
later merge of ground-plan into develop; section 7 lists them with the rest.

Reviewer: fresh adversarial agent. I changed no code, pushed nothing, posted nothing.

Blocking findings

None. I found no lost or doubled behaviour of either side, no broken CI job set, no wrong revision and
no hidden difference on the default path that the merge itself makes.

Non-blocking findings

N1. The plan bridge guard trips on the next Rust release or dependency bump, on unrelated develop PRs

  • rust-toolchain.toml says channel = "stable" with no version. The guard comment names stable 1.100.0.
    The plan bridge has 65 B gzip and 674 B raw headroom (builder/wasm/wasm_size_guard.py:29).
  • The step "Ground plan bridge size budget" runs in every wasm-bridges run, and web and web-browser
    need that job. So every develop PR that touches the web builds the plan bridge and checks its budget.
  • Scenario: the next stable Rust (6-week cadence: about 2026-12-24) or a bump of wasm-bindgen,
    serde_json or libm in Cargo.lock changes the code size by more than 65 B. Then each develop PR
    with web changes goes red, also when it touches no TrailView file.
  • Smallest fix, before the develop merge: a budget with a stated margin (for example the measured
    size plus the owner's "few per cent"), or pin the toolchain. The owner chooses.

N2. Develop loses every automated TrailView render in CI, not only the weekly job

  • The PR body names the lost weekly job (test-weekly.yml, tests: cut the slow TrailView tests and route the census by its inputs #3216). It does not say that the page-load
    test of smoke.test.js ("the page draws the fixture with no WebGPU error ... to 1 mm") also becomes
    @gpu (ground-plan 599b169). On develop that test runs on each PR on SwiftShader.
  • After the develop merge, no CI job loads and draws the TrailView page. @gpu runs only in the manual
    suite web.trailview-gpu on timosPC. still.test.js and home-views.test.js skip on SwiftShader.
  • This fits the owner's "TrailView tests lean" rule, but it is a coverage change for develop. Tell the
    owner in the develop PR body.

N3. wasm-bridges doubles in time and sits on the critical path of every web PR

  • wasm-bridges: develop push runs 2.2 and 2.4 min (43617472902, 43521754194); this PR 4.7 min
    (43618609569). The plan bridge build and "check 7" (57 s, "Every run") add about 2.3 min.
  • web-browser waits for it. So a develop PR that touches only the web goes from about 6.6 min to about
    9.2 min of wall time on that path. The PR body does not list this; it names only the test job.

N4. The data change of revision 20 on the default path was never shown to the owner

  • All default-path shots read the same data: the merged fixture tiles. The owner accepted the viewer
    changes on that data (2026-11-14). Develop's users get the viewer AND the new data after the next bake:
    new road and water heights of the solver, shallow tunnels with their tunnel tag (trailview features: a shallow tunnel keeps its tunnel tag on the ground #3382), culvert lifts,
    the baked stamps layer.
  • The develop fixture (25 tiles) differs from the merged fixture. A cheap proof: shoot develop on its own
    fixture against the merge on its own fixture, or ask the owner to accept the data change in words.
    The audit (4.6) asks for this acceptance; I found no record of it.

N5. The digest proof says "storeys off as in the bridge"; the bridge has storeys on

  • builder/ground-plan-wasm/Cargo.toml: default = ["storeys"]. The probe ran with --no-default-features
    (logs/dig2-merged.log). So the 256 digests prove the storeys-off build only.
  • The plan-path claim still holds: host/obc-ground-plan, builder/ground-plan-wasm, data/trailview, the
    fixture, plan/, roads/ and structures/ are byte-equal to ground-plan 46793ba (git diff --stat
    empty), and the bridge WASM is byte-equal. Correct the sentence in the body.

N6. sheets/default/attribution2.tsv is older than the PR table

N7. Small timing changes for develop's stills (not measured)

  • On develop the still page called TrailView.prepare(gpu) as soon as the device came, beside the
    release fetch. Now open() returns only after the release, so the compile starts after both. A few tens
    of milliseconds at most; one cold still.
  • On develop the other passes attached at the first drawn frame (or after 2 s). Now they attach when the
    view about the eye is final (or after 5 s), so a still compiles houses and instances after its eye view,
    not beside it. The device's compile cache helps from the second still on. I measured neither side.

N8. The test job: what costs, and when it runs

  • tools/ci/test.sh nextest-fast runs --workspace. Of the new targets only the census suite is carved.
    So every develop PR that runs test builds obc-ground-plan (with image, tiny-skia, spade,
    i_overlay) and runs its unit tests and fixture, storeys and transport, and the new
    obc-trailview-features census.
  • ci.ground-plan-census (faces, records, heights_fixture, mesh, graph) runs only when selected:
    its paths, or foundation = true, which is any change of Cargo.toml, Cargo.lock or the toolchain.
    Develop changed Cargo.lock in 6 commits since 2026-11-10, so the census will run often.
  • Measured: fast tier 174 s on develop push 43617472902 against 437 s on this PR (census selected). The
    cost of the always-on part (storeys and the rest without the census) is not measured.

1. The 9 conflict files (33897af)

  • engine.ts: sound.
    • start(canvas, gpu | Promise<Gpu>, config, view, scale, { plan, finest, pool, size }). All three
      callers pass the right shape: trailview.ts (plan options), still.ts ({ size }, a semantic edit
      in the merge commit), planner/trailview-start.ts (none). No caller passes a positional size.
    • output replaces develop's size field. size() returns output first, so ask() (frames before
      the device) and fit() use the still's size, not the size of the detached canvas. No other
      clientWidth read exists.
    • A still makes no Controls and no ResizeObserver (early return), as on develop; ground-plan added
      nothing after that point of the constructor.
    • context.configure with COPY_SRC and Targets(gpu, !!output) also in the re-attach after a lost
      device.
    • Frame loop: the grab resolves after the submit, then eyeDrawn when eyeFinal(). capture() runs
      after waitForIdle(), which waits for rest and complete, so a still never grabs before the late
      passes attach. A still is never in plan mode (plan defaults to false), so no plan worker is
      involved.
    • first(plan): prepare(gpu) on the still page and in the planner compiles the plan-off set, and the
      engines that follow are plan-off: the pipelines match.
  • trailview.ts: sound. open(plan) follows ground-plan's boot; main() passes the device promise;
    stills() awaits the device and calls prepare(). With no adapter, both pages show the notice through
    the final catch, as before. credits() gives the same list as ground-plan's inline code. See N7.
  • ci.yml: develop's job set plus ground-plan's steps (python.trailview in guards,
    OBC_SELECTED_SUITES, the plan bridge build, guard, check 7 and upload in wasm-bridges, the download
    in web and web-browser). No desktop job comes back. The other web builders: deploy-site.yml uses
    build-wasm-bridges.sh, which builds the plan bridge; no other workflow builds builder/web.
  • tools/justfile: the desktop recipe is gone; no file names builder/desktop or build:desktop.
  • vite.config.ts: develop's define and comments, ground-plan's plugins and worker plugin.
  • TrailView README: the three rows. terrain.rs: develop's FINE_ZOOM; z >= FINE_ZOOM equals
    z == 13 because the top zoom is 13. planner.rs: see 3. map.json: generated.

2. Semantic conflicts without a text conflict

  • I listed the files that differ from both parents (git diff-tree --cc). The merge commit edits
    still.ts (the options object), still-view.ts and covers/job.test.ts (plan: 'off'). The other
    files hold both sides' changes; the second and third merge commits add no edit of their own
    (terrain/terrain.ts and planner.rs show only the revision 20).
  • View.plan: the untyped JS views (still.test.js RIDGE, home-views.mjs) have no plan; the engine
    reads only plan === 'faces', so they draw as before. formatView would write plan=undefined for a
    view without the field, but no caller formats such a view.
  • harness.js: develop removed FLAT; no file uses it. shadows.test.js (removed by develop) was not
    changed by ground-plan.
  • resolve/targets.ts (develop's still argument), probeAt (ground-plan's kept argument), the
    stamp worker and the features decoder: no caller uses an old shape. features.ts FORMAT is 4 on
    both sides.
  • The new tags (levels, storey, slab) are read only by the bridge; no default-path TypeScript
    reads them.
  • ground plan GP12.3b part B: the viewer draws the decks, the device links the storeys #3482 (in 9464386) changes default-path shaders (corridor.wgsl, track-looks.wgsl, stamp.wgsl
    material mask 0xff to 15) behind PIECE or for codes over 15. The 12 default views equal to 33897af
    support "plan path only".
  • Develop's ce7d966 (one it.each title in roads/standing.test.ts): ground-plan never changed that
    file. No conflict in meaning. A scan of the web test files that ground-plan changed finds no other
    each title without a placeholder and no repeated title (the catalogue importer refuses doubles).

3. The features revision

  • History on each side since the base (revision 8): develop 9 to 12; ground-plan 9, 10, 13 to 19.
    20 is above both. Correct.
  • The revision is a key of obc data only; no tile or metadata holds it. The viewer opens the archive by
    features_format 4, the same on both sides (stream/terrain.ts:137). So the merged viewer reads
    released tiles of revision 12 by the code. The stamps layer is optional and checked by stamp_version.
  • The merged bake code equals ground-plan's but for FINE_ZOOM, so a bake of revision 20 gives the bytes
    of ground-plan's revision 19 (the fixture rebake gave no change). Revision 19 would have been correct
    too; 20 is safe.
  • No spec names the revision. specs/obc-data.md holds both sides' rows.

4. The proofs

5. What develop's users get

  • Bundle: the TrailView page entry +83 KB gzip (+30 %), the planner's lazy engine chunk +32 KB gzip.
    The plan-only modules in that chunk (plan.ts, pieces.ts, debug.ts, draw.ts, mesh.ts,
    inputs.ts, their WGSL and the bridge glue) are about 45 KB gzip as source, so a lazy import saves at
    most about a third of the growth. It is not a small change: Terrain uses Pieces, meshCells,
    stampNodes, CELLS_BYTES and COARSEST_MESH synchronously, and model.ts imports the glue. The plan
    path is to become the default. Fine to leave; tell the owner the figure.
  • CI: see N3 and N8.
  • WASM guard: see N1.

6. Tests and checks

  • CI run 43618609569 on 1b9bb6f: 26 pass, 2 skipped. The PR is MERGEABLE and CLEAN.
    origin/ground-plan is still 46793ba.
  • The one changed test file (covers/job.test.ts) only adds plan: 'off' to a fake view.
  • The author ran still.test.js and home-views.test.js (@gpu) on the merged build, not the whole
    web.trailview-gpu suite. The main page boot is exercised by the 191 + 8 shots.

7. Before the merge of ground-plan into develop

What must be true:

  • Merge develop again (now ce7d966 or later) and run CI green; take a new revision if develop bumps
    trailview-features again.
  • N1 settled (guard margin or toolchain pin), or the owner accepts red PRs at the next Rust release.
  • N4: the owner accepts the data change of the next bake in words, ideally with develop-own-fixture |
    merged-own-fixture shots.
  • The develop PR body names: the lost CI render coverage (N2), the wasm-bridges and test time (N3, N8),
    the bundle figures, revision 20 and the stamps step.
  • The Requirements: line for develop: I could not read the console (no agent token on this PC).

What the owner must know:

  • The PR has about 660 commits and one large review; CodeRabbit threads on develop are his to resolve,
    and the bot may be rate-limited on a diff of this size.
  • Revision 20 arrives with the next obc data bake; until then develop's viewer reads revision-12 tiles
    (by the code, not tested against a production release).
  • The next bake runs the new planner/trailview-stamps step (Node) and the new solver; TrailView features: the roads stage passes its memory budget with the new solver and ignores the cores it is given #3469 is open:
    a region bake needs taskset -c 0-3 today.
  • The owner's go for the push into develop is separate from any approval of this PR.

What I checked and found sound

The 9 conflict resolutions against both parents; the evil-merge files; the callers of TrailView.start;
the still path (size, controls, grab order, configure, targets, lost device); the boot of both pages; the
CI job set and the bridge downloads in every job and workflow that builds the web; test_plan.py and
suites.toml against develop; the revision history of both sides; the features format; the identity of
the plan code; two attributions from the PNGs; the re-merge diffs; ce7d966; the PR's Requirements:
line.

What I could not check

  • The time of the always-on obc-ground-plan tests (storeys, fixture, transport) in CI (N8): no
    clean CI data point, and I ran no test.
  • Still timings on develop against the merge (N7).
  • The verification console (obc req): no agent token here.
  • Which revision the production release holds, and the merged viewer against a real released archive.
  • The whole web.trailview-gpu suite on the merged build; I used no GPU.

Generated by Claude Code

@timohueser
timohueser merged commit 7fbbff9 into ground-plan Nov 15, 2026
28 checks passed
@timohueser
timohueser deleted the claude/gp-merge-develop-2 branch November 15, 2026 10:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant