Repository navigation
Description
Status: proposed
Part of #3105 (GP12, the crossing block of #3096). One pull request into ground-plan.
The steps of the block, in build order: GP12.1 #3342, GP12.2a #3343, GP12.2b #3352, GP12.3a #3344, GP12.3b #3353, GP12.4 #3345, GP12.5 #3346, GP12.6 #3347, GP12.7 #3348, GP12.8 #3349.
Read first: #3105 (words, invariants including the thickness rule, the gallery, the costs, the owner's decisions).
Design on timosPC: /media/timo/OSM1/tmp/ground-plan/crossings/CASES.md (case 2, 7.2, 7.3, 7.4), and above all
proto/LEARNED.md (sections 1 to 3a) and the approved prototype sheets proto/sheets/T02-*.jpg, T14-*.jpg,
A1-0-*.jpg. Short names: gp/ = host/obc-ground-plan/src/, tv/ = builder/web/src/lib/trailview/. Code
places verified on ground-plan ea4da43.
Goal
Near the eye, a tunnel way is a bore where it has cover and an open cut before it. The road runs in a cut until the
hill or the body of the way on top is high enough; there stands a plain concrete frame with a dark opening, and you
see into the tunnel. Nothing fills the opening, nothing stands out of the meadow, and the way on top is never cut.
An underpass under an embankment is the same construction: one plain box under the body of the way on top.
Depends on / blocks
- Depends on: GP12.1 (states BORE, OPEN, HIDDEN, portal lines, sheets of several ways), GP12.2a (the wall top),
GP12.2b (the ordinary end and the end face), GP12.3a and GP12.3b (the storey records, the storey mesh, the shell
corners, their drawing, the caster draw, the interim hand-over). - If its design note shows it too large for one review, it splits as GP12.3 did: a crate step (BORE storey,
Cover
face and bound, portal line, roof corners; no pixel change) and a viewer step (frame, end face with the opening,
holes off, bore look). - The places under towns (
R02,W5,R11) need GP12.8 part A and the box B rebake (at the owner's time). - Blocks: GP12.5 (the closed door and the pit build on the frame and the cut), GP12.6 (the far plate of a portal).
Design
1. Why ground fills the opening today (LEARNED 1, 2)
The cell mesh owns the ground about 12 m past a tunnel's tag end and draws its end band up to the terrain there. The
hole of the portal box (hole_distance, tv/terrain/terrain.wgsl:30-54, 79-80, 128; written by holes() in
tv/terrain/terrain.ts:1211, MAX_HOLES 32 at :49; boxes from portalBoxes in tv/structures/node.ts) exists only in
the terrain's pipeline, and the pieces never read it. The terrain mesh itself cannot make an opening: its level-18
cells are 1.6 m wide. So the cell mesh must own the cut, the end face and the ground over the bore behind the end
line; the terrain under the owned region is sunk already (plan_sunk, tv/plan/cells.wgsl:71-82).
2. Storey 0 at a bore (O4, O5)
- The open cut. From this step on,
classify(gp/graph/parts.rs:62-93) reads a tunnel tag on a ground kind
(carried by GP12.8 part A) asLevel::Tunnel.in_level(gp/faces/mod.rs:197-206) reads the tunnel states of GP12.1: an OPEN station is
on storey 0, a BORE station on its storey, a HIDDEN station on none. An OPEN stretch is a ground way of storey 0. Its sides follow the class rule: a bank where
there is room, a wall up towall_max_m(4.5) for a wall class, steeper above (owner: "B, the slope cut back as a
bank, walls only with no room"). A path class keeps its no-wall rule here; GP12.5 switches every OPEN stretch to
the no-room wall (owner: banks where there is room, walls only where there is none). - The portal line. The way's paved outline ends flat on the portal line of GP12.1; the line's edge kind is
Portal(exists ingp/faces/types.rs:51, unused). The cut's floor reaches the full width of the opening at the
portal line: where the way's own bands are narrower, the floor widens toward the line at the plan's
taper_rate(0.15 per metre; LEARNED 3.6). - The cover and its rank (LEARNED 3.1, 3.2, 3a.6). Behind the portal line, the ground over the bore's stroke
(its half-width plus its wall pluscover_side_m) is a new storey-0 faceCoverwith a floor bound: the ground
there is at least the bore's roof top plusroof_least_m. The bound is a rule of the heights in Rust and in the
WGSL port, as GP12.2b's ceiling. On free ground it may lift the terrain by at mostlift_max_m(the fallback of
GP12.1's state rule; owner: heights first). It never moves a paved face or a water face: under a paved way the
paved way is the cover and keeps its height.Covertakes its place inplan.jsonranksaftervergeand
beforeslope: under every paved surface, over every slope band. So it outranks every slope band of another way, so a neighbour's cut or
bank never enters the ground over a bore (the two tubes atT02lie 18.7 m apart). The bank of the cut runs round
the end against this raised ground, so the frame's sides are never free (the triangular walls of prototype
variant B), and the body of an upper way stays at its level up to the end face (the pale roof slab atA1). - The end face at the portal line (the headwall). The step on the portal line is GP12.2b's end face, with the
frame's outline left out of it: on the part of the line in front of the frame, the step's lower corners are the
floor's corners with the offset of the frame's top (the shell corner of GP12.3a). So the end face stands
only over and beside the frame, and no step closes the opening. Its look is GP12.2b's (concrete, a coping, no
pattern repeat, the smooth top of GP12.2a); it may be taller than 4.5 m. Where the ground over the face is the body
of a way, a railing of that way's class stands on the face's top. - No wall across a way's end. The portal end face replaces the old way end against the upper body: no step with
the stone look stands across a tunnel way at its portal line (the stone in today's arches atA1, LEARNED 3a.3). - Tunnel ends switch to GP12.2b's rule. The temporary end kind that kept tunnel ends as today goes.
3. The bore sheet and its shell (O2, O3)
- The floor. Storey −k of GP12.1:
records::near(.., −k, false)builds the floor of the bore's ways with the
strokes, ranks and overlay of storey 0. Ways of one sheet at different heights keep their own surfaces; the rank
decides where they overlap, and a riser stands between them (A1: carriageways and cycleways 0.8 m and 1.3 m
apart, LEARNED 3a.4). The storey mesh is GP12.3a's. - The shell. One cross-section per class: two walls, a flat roof and two cut corners (a class value; a path has
none). Built from the floor's free outline with corner offsets:
| Part | Construction | Thickness |
|---|---|---|
| Walls | from the floor's free outline up by the bore height (bore_m; on ladder step 2 down to bore_least_m) |
0.6 m (today's box walls, WALL_M of tv/structures/tunnels.ts) |
| Roof | the floor again, at the lower of floor + bore height and the ground over it less roof_least_m |
at least roof_least_m (0.3 m) of slab between the bore and storey 0 |
| Frame | a ring about the opening on the portal line: jambs, lintel, reveal, coping | jambs 0.7 m wide, lintel 1.3 m tall (CASES case 2), 0.6 m deep along the way (today's FACADE_M), coping on top |
| Plain end wall | where the sheet ends inside the ground (a closed line, the data's end) | the wall thickness, closed |
- A roof corner has two labels (GP12.3a's shell corner): the floor's with the bore height, and the storey-0 face's
over it lessroof_least_m; its height is the lower one. So the GPU gives the roof the drawn ground less the
plate where the ground is low: no roof shows over the ground, whatever the heights. - The frame and the sheet share the portal line and its row of points (LEARNED 3.4: a gap of 0.4 m showed ground
inside the bore when they did not). - Side-by-side bores whose strokes touch are one sheet with one frame (
A1: one box, not three arches). Bores
with a gap have a frame each, on its own line (T02: two frames 3 m apart along the way read right).
- Inside the bore. The pieces draw the bore (it is cell mesh, not a structure mesh). The floor has the way's look;
tv/plan/piece.wgslgets the bore material for the walls and the roof: a port of theBORE_FACElook of
tv/structures/structures.wgsl(:68,:269: tiles, lamps). Dark inside, as today's bore.
4. What stops on the plan path, and the interim far look
nodeStamps(tv/terrain/bins.ts:1048) writes no record for aTUNNELline withplan: the tunnel terms of the
stamp (COVERfor a tunnel,PORTAL_CORE,PORTAL_CELLS,COVER_M,BORE_WALL_M,COVER_REACH_M:
tv/terrain/stamp-core.wgsl:92, 160-166, 686-692, 746) get no input there. The gallery's use of theCOVERbucket
(:734-739) stays until GP12.7.- The terrain's portal holes are off on the plan path (
HOLES,holes(),portalBoxes). Inside the mesh range the
plan owns the ground; past it the old facade's dark void face covers the opening until GP12.6. tunnels()(tv/structures/tunnels.ts:132: bore, facade, box) stays the far look until GP12.6, with GP12.3b's
mechanism: the structure pipeline discards its fragments whereplan_meshedholds (the cell's mesh draws, so its
bores draw). This hand-over goes in GP12.6.- The edge kind
Headwall(exists ingp/faces/types.rs:50, unused) stays for the culvert mouths of TrailView ground plan GP15: water crossings: culverts, fords, pipes in stones, stepping stones, footbridges #3108; the
portal end face usesPortal. - The default path keeps all of it until GP19.
Data and bytes
- Records: the storey −k lines and candidates (estimate under 2 kB per 20 m underpass); the
Coverface's items in
the bins (estimate 0.3 kB per 20 m of bore). - Mesh: the bore's storey triangulation and shell (CASES 7.6 estimate: about 170 triangles for a 20 m underpass with
two frames, 150 per 100 m of tunnel, 30 per frame). - Values in the
structuresblock ofplan.json:cover_side_m(the margin of the cover past the bore's wall;
1.0, as today'sBORE_WALL_M), the frame (frame_side_m0.7,frame_top_m1.3,frame_depth_m0.6), the cut
corner per class.
Invariants
- No terrain or ground face lies inside an opening: from the portal line into the bore, between the floor and the
roof, nothing of storey 0 draws. Check: a crate test over the fixture's portals (no storey-0 triangle over the
bore's floor under its roof within 12 m of a portal line), and the gallery. - The way on top is never cut: no storey-0 paved face over a BORE station changes its outline or its height. Check:
the faces and heights of the upper ways equal those of the base over the fixture, except theCoverface beside
them; andA1's tracks in the gallery. - The roof lies at least
roof_least_munder storey 0 at every roof corner. Check: crate test over the fixture's
bores with the Rust heights. - A frame and its bore share one end line and its row of points. Check: crate test.
- Every shell part has its stated thickness and closes on every visible side; no bore wall stands free over the
ground behind a frame. Check: the shell closure test of GP12.3a, and a test that every bore wall's outer face lies
under storey 0 behind the portal line. - No ground of another way enters the
Coverface. Check: faces test over the fixture. - No stone-look step across a way at its portal line. Check: a count over the fixture's portal lines, target 0.
Decisions already made by the owner that apply
- "Tunnel mouth, road and path: B, the slope cut back as a bank, walls only with no room" ("go with B, I agree").
- "Ground lifted up to 1 m over a shallow tunnel: okay, but isn't the real solution to make the tunnel sit lower?" →
heights first; the lift is only the fallback. - "End walls and tunnel frames taller than 4.5 m: yes of course."
- "The tunnels cut through the rails and make them look pretty bad": the rails stay whole above an underpass.
- Attinghausen, today's heights with the new drawing: the underpass reads right to him; he asked about "the wall
going straight through it" in the view from the east: find what fills that opening before this step claims it. - Structure parts are never paper-thin sheets: real thickness and closed faces.
Not in scope
The closed framed door, HIDDEN stretches as a look, and the path subway's way down with walls (GP12.5). The far plate
(GP12.6). Galleries (GP12.7). Rails in a bore (#3106). Lamps, portal buildings, ventilation, the look of the frame
beyond plain concrete.
Proofs and pictures required before review
- The gallery, all 50 places, before and after, on the demo release, one before | after sheet per place with label
boxes, a verdict table besidegallery/INDEX.md; the implementer opens every sheet.- Must turn BROKEN to GOOD or IMPERFECT (full or thin room on today's data):
A1,N5,N6,T02,p04,T07,
T14,T09(thin),v2(its two tunnels in the bank). - Must improve:
T01(frames at the cover's line, no wedge of wall on the carriageway),T05(no free plate
from the lake side, no box),T06(no hillside in the opening),R01,S3(no box on the approach road),
R11(no box walls over the junction). - Must stay GOOD:
B1; and every GOOD deck place. - Places on step 3 or inverted today (
S1,S2,S7,S8,W3,W4,R03,R08,R02,W5) end whole
with an ordinary way end here; their door is GP12.5's.
- Must turn BROKEN to GOOD or IMPERFECT (full or thin room on today's data):
- The
A1view from the east (cameraeastofproto/views.json): a probe and a picture that name what stands in
the opening (the owner's question). If it is the approach behind the box, show it from two more heights. - The prototype's comparisons again on this build:
T02(eye,drive,near,obl,side,top),T14,
A1(proto/views.json), against the approved prototype sheets. - Far views of
B1andT02from 600 m and 1.2 km: the interim far look (old facade and void past the mesh range)
has no ground in its opening and no doubled frame at the hand-over. - The owner sees the sheets and gives his OK before the merge.
Cost to measure
Base: ground-plan at the branch point, and today's tunnel meshes for the same tunnels.
- Triangles and vertices per bore and per frame against today's
tunnels.tsmesh (bore, facade, box) of the same
tunnel, on the GP12 design sample (gp12-design/measure/gp12-share.ts). - Cell build time native p50 and p95 over the fixture cells with a tunnel and over all cells; WASM time on
plan-timing.html. - Record and mesh bytes per cell (sum, p50, p95, max).
- Frame:
walk,valley-driveandluz, against the base and develop's head. The terrain without the hole pipeline
on the plan path: report the terrain pass. - WASM bytes, gzip and raw (the owner accepts about 15 to 25 kB for the block; report the step's share; stop past
25 kB in total).
Risks and the stop rule
- Stop and report if the opening needs a pixel discard in the terrain or in the pieces (a discard turns off the early
depth test of tile GPUs: the phone). - Stop and report if a mouth, an underpass or a frame needs a shape of its own, a repair pass, or a case distinction
by place or by class pair; or if theCoverrank breaks a way's own faces. - Stop and report if any cost reaches twice its base.
- After two failed review rounds, stop and report.
Open points
- Catenary masts over a box (LEARNED 3a.8): a mast whose foot lies on the cover or on the fill behind an end
face sinks its foot into the box roof. The masts are TrailView ground plan GP17: point objects: street lamps, OSM point kinds, catenary masts, aerialway pylons and stations #3116's (they slide along their line in the bake). Count the
masts within a bore's stroke on box B and report; do not move masts here.
Requirements: the verification console could not be checked from this PC (no agent token here). The orchestrator sets
the pull request's Requirements: line.
Deviations of step 1 (pull request #3488, the levels rules in the bake)
classifydoes not readtunnel=yeson the ground kinds in step 1. On the device,classifygives each part
its level. A shallow tunnel would become a tunnel way,level_tags::splitkeeps a tunnel way's level, and storey 0
draws only ground ways: the shallow tunnels would go from the plan (fixture: 6 features, 11 copies; the census
tiles of box B: 191 copies). The reading moves to step 2, together with the split of tunnel ways by their states,
so that the OPEN stretches of a shallow tunnel stay on the ground. Measured: with the reading alone, 11 of the 256
fixture cells change what the viewer reads. 11 census portals have a shallow tunnel at their
end; their counts can move in step 2.- Rule B does not change the lift. Only a stroke more than the at-grade band over the floor hides a station, but
every paved stroke still forbids the 1 m lift on free ground (a paved face never moves). The census effect is that
of the diagnosis's simulation. - Census: 20 closed mouths, not 19 (base 55 on the terrain that the bake reads; the diagnosis's 58 read
zoom-12 terrain only). Three mouths stay closed for a cause outside rules A and B, so the levels see no tag end at
their portal:- P0134, P0215: the tunnel way runs through a junction farther in. The node map of the levels (
Ctx::nodesin
graph/levels.rs, one entry per way and direction) takes that junction for the way's end node, so the tunnel
seems to go on at its portal. General fix (not built): enter an arm in the map only at its way's end station.
Measured on box B (not in the pull request): it opens P0134 and P0215 and closes none, but it also changes the
levels of 6 deck features, so it needs its own proof on bridges. - P0205: two tunnel ways and a bridge meet at one node at the portal, so a tunnel goes on there. General form (not
built): a way end is a tag end where a way of another kind (not water) meets it.
P0433 and P0797 (counted as heights by the diagnosis) open by rule B.
- P0134, P0215: the tunnel way runs through a junction farther in. The node map of the levels (
Activity
timohueser commentedon Nov 11, 2026
From GP12.1 (PR #3367): what this step can read, and what differs from the plan
graph.levels.runscovers every tagged way (non-synthetic deck way, non-water tunnel way) once, in order; untagged ways that end on step 3 also have runs (HIDDEN under the body, OPEN elsewhere).in_levelis unchanged.graph.levels.end_lines(A, Portal, Closed); merged lines list each member way with its station on the merged line, and the members' runs are moved to that station.edge_band_min GP12.1's sheets (the block has no wall value): GP12.4 should add one tostructures.deck(upper)of an untagged upper way is itsdeck_least_m.A1east,p04). The rule "the floor is at or under the terrain" may need a tolerance before doors are drawn (GP12.4, GP12.5).Generated by Claude Code
timohueser commentedon Nov 12, 2026
Orchestrator note after the merge of GP12.1 (#3367). Earlier notes of the implementer on this issue are stale in three points: the end-line merge, the sheets and the storeys are NOT in the merged step (their code is on the branch
claude/gp12-1-later, to arrive with the step that first uses them); the floor-over-terrain tolerance is fixed; the levels stage is not linked into the device build (GP12.3a links it, target at most 12 kB gzip).Open points from the delta review (
crossings/REVIEW-3367-delta.md) that the step which first DRAWS from the states must close before it draws:graph/levels.rsabout 393-405): a way that joins at a tunnel end is left out of the ground along the whole tunnel, not only near that end. A spiral tunnel or a hairpin over a portal can lose its cover or show OPEN under a road. Limit the exclusion to the reach of the tag end.A_MAX_M= 40 m has no stated source, and past it the whole deck way becomes ground;GRID_Mholds an unexplained 13.0; the real read reaches 42 m against the 40 m of the reach row.Generated by Claude Code
timohueser commentedon Nov 12, 2026
Correction after the merge of GP12.8 part A (#3382), from its proofs and its review (
crossings/REVIEW-3382.md).The body of this issue says or implies that part A (the bake keeps the
tunneltag on a tunnel that the shallow rule turns into a ground way) unblocks the gallery places R02 (Kriens), W5 and R11. That is wrong. The motorway tunnels at those three places are of the kindtunnelin every bake (the demo data, round 1 and the new bake); the shallow rule never touched them. Their faults come from the road heights of the demo data and from the old tunnel drawing. On round-1 heights R02 is whole (crossings/gallery/r1-vs-today/).What part A does give: on the test box 194 shallow tunnels (mostly short tunnels under buildings) keep their tag, as
tunnel=yeson a ground kind, so that the step that draws tunnels can see them. No reader reads that tag today.For the step that first reads the tag (#3345): tiles of feature revision 15 that were baked before #3382 do not carry it. That step must depend on a rebake (the revision 16 of #3337 or a later one) or raise the revision itself; otherwise regions on older tiles silently keep their shallow tunnels as ground ways.
Also wrong in the body of #3349: "or the mapped value" (the tag's value is always
yes) and that the ground kinds' keep lists includetunnel(they do not; the tag is written by the shallow rule itself).Generated by Claude Code
timohueser commentedon Nov 12, 2026
From GP12.2b part 1 (#3352, proof in
crossings/gp122b/PROOF.md): the stamp's tunnel termCOVER(tv/terrain/stamp-core.wgsl:102, 164, 690) lifts the ground beside the roads atR02(+10.2 m against the crate's heights) andW5(+4.8 m), and adds height atR11on round 1. With the plan'sheight_ofoff the stamp, the walls there go (29 % and 16 to 19 % of the pictures change). GP12.2b part 2 does not touchCOVER: stopping it on the plan path is this step's job (section 4 here), and the end faces at those three places are judged here, not in GP12.2b.The proof also gives this step two inputs: the portal's end face needs GP12.2b's ceiling in its bore form (the ground over the bore held under the cover), and the square end with its band fade (
gp/faces/heights/mod.rs:1322) makes a 12 m face insideT01's frame once the stamp is off.Generated by Claude Code
timohueser commentedon Nov 12, 2026
For this step, from the rebake of the test box (see the correction on #3384): the bake's floating rule for SHORT tunnel parts (no sample farther than 20 m from both ends, any sample more than 10 m over the terrain) judges each piece between two way splits. It therefore drops way 1419883678, a 37.7 m piece of the A2 Sonnenberg tunnel chain at its south portal. When this step draws bores from the tunnel lines, that piece is a gap at a portal. Repair it by a general rule in the bake before this step relies on the lines: the length that decides "short" is the length of the joined tunnel chain (pieces of the same way class and tunnel tag that share end vertices), not of one piece.
Generated by Claude Code
timohueser commentedon Nov 12, 2026
From GP12.2b part 2 (PR #3396), for GP12.4
heights::ceiling, WGSLh_ceiling) bounds the terrain and slope rules only; a bore form can reusedecks()with the sign turned. The stroke's flat ends are stored per ceiling line (words 4 and 5).nodeStampsnow writes no record for aBRIDGEline on the near levels withplan;TUNNELlines keep theirs (COVER, portals) for this step.Heldstep over the cap is labelledHeldOverCapand listed bycheck::held_over_cap.Generated by Claude Code
timohueser commentedon Nov 12, 2026
From the review of PR #3396 (GP12.2b part 2): a point for this step
R11do not come from the ceiling (mincannot lift). With the plan,nodeStamps(tv/terrain/bins.ts) writes noBRIDGErecord on the near levels, so the old trough's wide clearing is gone and the DEM returns beside the stroke. The DEM hill over the junction then goes through the old tunnel boxes (old structure code), and the deeper cut under the slab gives the stone walls. This step replaces the tunnel boxes and stopsCOVER; judgeR11after it.Generated by Claude Code
timohueser commentedon Nov 13, 2026
From GP12.3a (PR #3464): what this step can read, and what differs from the plan
faces::deckselects the ways of a storey over the ground fromgraph::levels(DECK runs; synthetic decks whole); storeys under the ground (bores) are not built:Decks::oftakes decks only, andmesh::shellbuilds deck parts only. A bore storey needs its own selection (BORE runs, storey −k) and the shell's sign turned.STOREY, storey in the leg bits); a reader of a storey's label reads only its storey's lines (heights::present).storeysis off; with it the bridge is +62.1 kB gzip (levels included), which passes the block's 15 to 25 kB already before the bores.Generated by Claude Code
timohueser commentedon Nov 13, 2026
From GP12.3b part A (PR #3470), for GP12.4
The levels of the way graph (
graph::levels) now run in the feature bake (features revision 18), on every zoom-16 cell of each zoom-13 tile with a road way. The zoom-13 tiles carry per road way the tagslevels(the states D/B/O/H and the end lines A/P/C, in metres along the stroke),storeyandslab(specs/trailview-features.md, Structures). The device reads them withgraph::level_tags::read; it does not rungraph::levels. The bake uses the rule tables of its own build: a change of a level rule (cover,lift_max_m,open_max_m, the least rooms, the class rows) or of the code of the levels takes a new bake, a new revision and a fixture rebake (make-fixture.py features, about 10 s); the testthe_tiles_carry_the_levels_of_each_cellfails until then. A way of a cell margin reads the levels of the cell that owns it, so cells agree by construction. The low zooms carry no level tags.For this step: the BORE, OPEN and HIDDEN runs of tunnel ways and their portal and closed lines are already in the tags. Tuning the bore rules from renders means a rebake per round.
Generated by Claude Code
timohueser commentedon Nov 13, 2026
Diagnosis for this step from the owner's notes on checkpoint 5 (
ground-plan768d39e), 2026-11-13. The owner: "We still get meadow and other stuff in tunnels everywhere, this needs one really solid fix … it's not just this tunnel, it's most really", and "all the tunnels have wall etc. covering their entrance, this cannot stay this way, it's one of the most noticeable flaws we still have."This comment is part of the contract of GP12.4. The build must remove the classes G, Wa, Wb AND Wc below by one general rule each, and must show it on the whole census (1,014 portals of test box B), not on single views. The owner sees the cause and the rule before the build. The full report, the probes and the sheets are on timosPC in
/media/timo/OSM1/tmp/ground-plan/cp5-notes/(REPORT.md,sheets/p1-class-*.jpg).Also binding from the WASM review of #3470: the crossing block holds at about 60 kB gzipped on the device only if bores and portals REUSE the storey records and the shell of the deck steps and add about 8 kB or less.
Point 1: tunnel mouths. Census of box B (1,014 portals; 1,013 with a floor height)
Every portal is sorted by what the crate draws on the tunnel axis, then checked against eye shots: 57 portals
shot at eye height from the road, 30 m out; 46 show the mouth. Data and shots agree on all 11 wall shots and on
22 of 28 ground shots. Data and script:
logs/portals.json,logs/census.json,scripts/classes.py.tv/terrain/bins.ts:1100drops roads, paths, rails and near bridges, not tunnels). PORTAL_CORE holds the floor for about 3.3 m (stamp-core.wgsl:163,746), then COVER lifts the ground to floor + bore + 1 m (:164,686-691). The plan's ground corners read that page (plan/corners.wgsl:70-85). The plan mesh owns the bore's first metres (R = sharp cells grown by FACE_RING 2,mesh/params.rs:22; 992 of 1,014 portals), but the pieces never read the terrain's portal holes (terrain.wgsl:37-54,128). Near the eye the facade's void is dropped (structures.wgsl:93,262), so the eye meets the ramp.Held/HeldOverCap/Unrelated,faces/heights/mod.rs:1700-1729).LEVELend (records/mod.rs:880-893) lets the way's own cut cone run past its end. Wall classes get a wall up to 4.5 m across the end (P0734, Axenstrasse).Also at mouths: 109 portals lie on flat ground (frame and box stand out of the meadow); at 27 portals the floor
is more than 1 m over the terrain; trees or plants stand in 5 of the 46 openings seen; P0870 is open only because
its cell draws no plan mesh (
debug=ownedis all red there; cause not found).Evidence for G: drawn terrain on the axis of 42 G portals (
trailView.check), median height over the floor:0.0 m at 2 m inside, 2.1 m at 3 m, 3.9 m at 3.5 m, 5.6 m at 4 m, 6.2 m at 5 m, 7.5 m at 8 m.
debug=owned: theplan owns the bore at P0214, P0076 and P0734 (all filled) and not at P0202 (open).
Does GP12.4 (#3345) remove the classes?
Cover face, the end face with its opening, invariant 1): 502 of 579.
section 2.5 bans a stone step across a way end (invariant 7).
GP12.5 and will still read as "covered".
(its stroke between floor and roof, behind the portal line) outranks every storey-0 face, paved ones included;
a way whose paved face reaches into it ends at the opening's side.
cover's reach), OSM walls draped over portals, and the far look until GP12.6.
An early general fix
One small rule removes most of the filled mouths: the plan pieces read the terrain's portal holes (a
piece.wgslvariant with the same
hole_distancetest asterrain.wgsl:128, used only for cells that hold a portal box).Cost: about 40–60 lines of WGSL/TS, 0 WASM bytes, one more draw variant; the pieces of those cells lose the early
depth test on tile GPUs. Estimated effect: about 940 of 993 filled mouths near the eye. Not built, not measured.
It breaks GP12.4's stop rule ("no pixel discard in the pieces"), so it is the owner's decision; GP12.4 would
delete it. Two other options do not work: handing the bore cells back to the terrain in the region R (works only
per 3.3 m cell, does nothing for Wa/Wb/Wc, breaks the 5×5 shape the strip relies on); stopping the stamp's tunnel
terms alone (the hill still fills the opening at 904 portals and the end steps grow to up to 7 m).
Generated by Claude Code
timohueser commentedon Nov 13, 2026
Owner decision, 2026-11-13: "Wait for the full step". The early fix (the plan pieces read the terrain's portal holes) is not built. Tunnel mouths are fixed by GP12.4 itself, as one general rule for each class of the census above.
Generated by Claude Code
timohueser commentedon Nov 14, 2026
GP12.3b part B (#3482) changes what this step builds on:
graph::build_levelled,level_tags::split): each bridge way is splitat the end lines that the tiles carry, and each stretch takes the level of its state. Tunnel ways are not split:
splitskips a part whose level isTunnel, so a tunnel way keeps today's level. This step can split them by thesame function (a BORE stretch on the tunnel level, an OPEN stretch on the ground).
Generated by Claude Code
timohueser commentedon Nov 15, 2026
Proposal: the updated contract of this step, after the tunnel diagnosis (waits for the owner's answers)
Status: proposal. It waits for the owner's answers to the four questions at the end. Diagnosis, probe, sheets and
drawings on timosPC:
/media/timo/OSM1/tmp/ground-plan/tunnels/(DIAGNOSIS.md,sheets/INDEX.md). Code checked onground-plan46793ba. Nothing is built.1. The closed mouths, counted again
The census comment above gives "about 89" closed mouths. That count read the end lines of every tunnel way within
60 m of a portal (twin tubes, the other end of a short tunnel). A new probe reads each portal's own tunnel way from
its tag end. Result on the census (1,014 portals of box B): 955 mouths get a portal line; 58 are closed (54 hidden at
the tag end, 4 with a closed line less than 5 m behind the portal line); 1 has no way end in the data.
Expected with this step: 994 open, 19 closed (15 wait for #3349), 1 without a way end.
2. What holds, and what changes in the design
(the cover face, its rank, its floor bound as GP12.2b's ceiling with the sign turned); 2.4 (the end face with the
frame left out, built as a shell wall between two labels, no wall cap on the edge kind
Portal); 2.5; 2.6 (thelevel end goes; a portal end is the square end of a stub, end kind
DECK, with the bank round it against thecover); 3.3; invariants 1 to 7.
in_level.level_tags::splitstops skipping tunnel ways: a BORE stretch takes the tunnel level, an OPEN stretch the ground,a HIDDEN stretch goes.
classifyreadstunnel=yeson a ground kind as a tunnel (the bake then gives these wayslevels too).
ground_junctionsreads "deck or bore" in place of "deck".faces::deck::Decks::ofgives a bore way storey −1;records::storeysbuilds every storey but 0, with a signed storey in the leg bits.mesh::shell::mesh_storeyswith the sign turned: the top is the floor, theunderside at the bore height is the roof, the edge faces are the walls (turned to face in), a closed end gets the
plain end wall. No parapet, railing, post or pier below storey 0. New parts: the frame (
prismon each portalline) and the end face. The roof corner's second label is new in the corner format.
GP12.2b; the far levels keep them until GP12.6. The terrain's portal holes stay until GP12.6: they act only where
the terrain draws, past the mesh range. The old tunnel mesh is discarded where the plan draws, by the bridges'
mechanism (
plan_share).open_max_mof a tag end, a tunnel station with too little coveris OPEN, whatever the height of its floor over the terrain model. The way goes on as a ground way on its bank until
the cover holds; there is the portal line. This is the general form of TrailView ground plan GP12: bridges, tunnels, underpasses and galleries: decks, bores and roofs with the ladder (the contract of the crossing block) #3105 open point 1.
surface lies more than the snap height over the floor.
that does not cross over the bore, paved ones included. Such a way ends at the side of the opening with an ordinary
end. A way on top always has a crossing with the bore, so it is never cut.
a bore's stroke near a portal line; 10, the census counts of section 4 below.
3. Size
The bores reuse the tags, the split, the deck storeys, the storey records and the shell. Estimate on the device:
about 4 to 7 kB gzip (graph 0.2, faces and records 0.6, shell 0.6, the roof's second label 1.0, frame and end face
1.5, cover face and bound 1.5, opening rule 0.3). Not measured. The chain is at about 49.4 kB of about 60 kB; this
leaves about 4 to 6 kB for GP12.6 and GP12.7. Past 10 kB, the first cut is the roof's second label (the shader reads
storey 0 at the roof corner, about −1 kB), then the frame's coping and reveal (about −0.4 kB). Rules A and B are bake
code: 0 device bytes.
4. The build order and the proof on the whole census
a new features revision, the fixture rebaked, box B rebaked by the step's script. Proof: the probe over all 1,014
portals, closed mouths by reason before and after (58 to 19); the levels test of the fixture; the tiles differ
only in the
levelstags of tunnel ways.(floor, walls, roof, frame, end face, end wall), the cover face and its bound in Rust, the opening rule, the portal
end kind. Proof, crate counts over all census portals of box B, each to 0: storey-0 triangles between floor and
roof within 12 m of a portal line (class G); stone-look steps on a portal line (Wa, Wb); storey-0 faces of a way
without a crossing over the bore in an opening (Wc); roof corners less than
roof_least_munder storey 0; shelledges that do not close. Debug pictures of the storeys at T02, A1, T14. WASM bytes from CI.
levels, the old near tunnel mesh discarded, the plants rule. Proof: the gallery's 50 places before and after, the
57 eye cameras of the census shot again, the 12 closed mouths of the diagnosis, far views of B1 and T02 from 600 m
and 1.2 km, the GPU probe of the drawn height on the axis of the 46 mouths that the census saw (nothing between
floor and roof from 2 to 8 m inside), the frame cost on
walk,valley-driveandluz.GP12.5 then draws the doors at the 19 closed mouths; #3349 takes the 15 height faults as its test list.
5. Risks
The bank round the end against the cover (torn faces); the roof's two labels on the GPU and a roof over low ground on
ladder step 2; the end face beside the frame under a paved way on top; rule A at the 8 mouths after a bridge (a bank
of up to 5 m between the bridge's end and the portal); the opening rule where the graph misses a crossing; the
hand-over at the mesh range until GP12.6; the size estimate (the deck steps came out 3 times over their first
estimate).
6. The owner's answers that this proposal waits for
deck to the portal; (c) a door.
a cut (heights); (b) a light cover at ground level (a box on the ground, as GP12.7).
Generated by Claude Code
timohueser commentedon Nov 15, 2026
Owner, 2026-11-15, on the proposal above: "I guess you can merge step 1". So pull request 1 (the two levels rules in the bake, no pixel change) may merge after its review and green CI.
The other three questions of the proposal were not clear to the owner, and they are technical defaults. The orchestrator takes the recommended option for each, and the owner judges the real pictures of step 3:
If a result looks wrong in the real pictures, it changes then.
Generated by Claude Code
timohueser commentedon Nov 15, 2026
Notes for step 2, from the review of step 1 (#3488)
The review of #3488 (rules A and B in the bake's levels) found no blocking fault. Three of its notes are for step 2 and the steps after it:
tunnel_statereads the paved stroke at the tunnel station's middle only (ground(a.p, w));ladderreads the least of five corner pairs. Both use the same band (the lowersnap_height_mof the two ways). On a banked or sloped way within a few centimetres of the band, the ladder can say at grade while the station is HIDDEN, or the reverse. No action in step 1. In step 2 both decide what draws, so make them read the same points, or show that the difference cannot be seen.classifyland together (review N8). Whenlevel_tags::splitstops skipping tunnel parts, the fixture testthe_tiles_carry_the_levels_of_each_cell(tests/graph.rs) must stop skipping them too (if part.level == Level::Tunnel ... continue).classifymust readtunnel=yeson the ground kinds in the same pull request as the split: alone, it removes the shallow tunnels from the plan (11 of the 256 fixture cells change), and no test guards against it.Review:
/media/timo/OSM1/tmp/ground-plan/tunnels/step1/REVIEW-3488.md.Generated by Claude Code
timohueser commentedon Nov 15, 2026
Note for tunnel step 2 (from the review of #3489): on tiles of an older bake a way through a building is now
Level::Groundbut still carriestunnel=building_passage. The bake keepstunnel=yeson the shallow tunnels that it turns into ground ways. So step 2 must testtunnel == "yes"(or the way kind), not!= "no", when it reads the tag on ground kinds; else it takes a building passage for a shallow tunnel. The water branch and the debug tool use!= "no": do not copy that test.Generated by Claude Code
timohueser commentedon Nov 16, 2026
Step 2, redesigned (design revision 2, after two failed build rounds and two adversarial reviews). This comment replaces the step 2 part of the earlier design in this issue. The build of the crate body starts from it on the branch
claude/gp12-4-bore-body.Open points from the second review, which the build and the later bake step must close. They are not solved by the text below:
decks()changes decks with the feature off. It becomes its own slice.GP12.4 step 2, redesign: a tunnel is a lid over a way (round 2 of the design)
Status: proposal of the designer, after the review
REVIEW-DESIGN.md. The first design isDESIGN-r1.md. Theowner's tunnel decisions and the contract of #3345 stand. A spike proves the foundation and measures the mouth
(local branch
claude/gp12-4-joint-spike, 8938a7348; not for merge, not pushed). All numbers are in "Evidence".Summary
A tunnel is no second ground. It is a lid over a way. Storey 0 and the floors under it are two labellings of
one map. The tunnel body is the prism between a floor and its lid, on storey 0's own triangles. The review
confirms this part: the joint, the closure and the roof under the lid are 0 on all 1,014 portals.
The review shows that the first design did not measure the owner's top flaw: an entrance that is filled or covered.
This round adds one count for it, the gate: an entrance that a rider cannot see into. The same probe measures it
with the bores off, on round 2, on the first spike and on this spike.
Of the 1,014 entrances of test box B, the gate holds 1,014 with the bores off (what
developshows: noentrance has a body), 571 on round 2, 192 on the first spike and 104 on this spike (10.3 %). 30 of the
104 have one fault only: a wall in front of the mouth that stands there with no bores too.
Ten general rules bring the count down. Each one is a cause that the spike found, not a tuned value. The count is
not near 0. The largest part that stays is a wall in front of the mouth that stands there with no bores too: storey
0's own rule for two ground ways. "The entering way owns its approach" was tried in two forms; each one made more
walls over the box than it removed, and both are off. "The list that stays" names each part and its cause, and "The
fallback" says what to ship.
The size is over the target. The spike without its test switches adds 23.7 kB gzip. The estimate for a written
build is about 15 kB (13 to 17), and about 11 kB (10 to 13) with two decisions moved into the bake (section F).
Words
as far as
open_max_malong them.line, under a lid (the last metres of the entering way).
bore_least_m + roof_least_mof the lower way's class: 2.2 m for a path, 3.3 m for a road, 5.1 mfor a railway. The bake's BORE test uses the same number.
floor, at every place where their strokes meet. Its lid is its carriageway and its shoulders. Its verge is ground.
The foundation: six rules
level: OPEN is a ground way, BORE is a bore, HIDDEN is nothing.
its offsets and a corner at each layer of the two ways and at the bore's wall. The floor bound of the heights ends
flat on the same line.
can have a floor label too. Storey 0's faces split where the floor label changes.
other face yields. The cover's height is the ground's height, held at the least room over the floor. No cover
lies on an entering way's open stroke where no floor lies under it.
points. A roof corner is the lower of the floor plus the bore's height, and of each lid corner at that point over
the same floor less the roof's plate. A floor corner has the same caps. Each edge of the floor's outline has a
wall. An edge of the opening has no wall: it has the front of the lid's plate over it and a sill slot under it.
body_mbehind each of its portallines. The rest is HIDDEN. A bore shorter than twice
body_mis whole.Rules 1 to 5 are in the spike. Rule 6 is in the spike as a test switch that reads the bake's tags; the bake's rule
itself is not built.
The gate: an entrance that a rider cannot see into
The measure (probe
probe/census4.rs, all 1,014 census portals of box B, the 3 x 3 cells about each).A portal's own opening pieces are storey 0's portal edges in the corridor of its way: from 6 m in front of the
census point (the tunnel's tag end) to 45 m behind it, 8 m beside it and a quarter of the metres behind more, less
the pieces that look against the portal (the far end of a short tunnel). A portal is in the gate when one part
holds:
bore_least_mof the floor's own class less 5 cm. A paved face or a shard in the opening, and a roof at or under the floor, are low samples.frame_depth_m).The table also gives two wider views that are no part of the gate: "low strict" (one low sample is enough) and "wall
within 3 m". With the bores off the pieces are those of the spike's run, so "wall" and "step" measure the same
places.
The numbers (portals).
developshows)pieces are the spike's).
36 it stands with and without; at 11 only the bores make it.
section F, the sizes and the digests were measured on it.
portal in the next feature is hidden as a whole: "none" rises by 37. That is the switch, not the rule; the bake's
rule walks the chain. The other parts do not move with the body's length (low 39 to 36, wall 69 to 67).
probe/census4r2.rs, the same tests).By class (the census class of each portal; this spike).
By kind and by the tunnel part's length (this spike).
How open the low mouths are. A low mouth is not always a closed one. The probe also takes, per portal, the least
mean of
min(room, least) / leastover the 0.5 m bands behind the pieces.What fixed which part
Each row is one cause, found at one portal and then measured over the box. No row holds a value that was tuned.
lift_max_m. The device's terrain is short at 1 to 2 m behind many portal lines: the bake tests stations, and it puts the line between them.Two rules that were tried and are off (both are "the entering way owns its approach"; switches in the spike; the
three gate numbers are runs of one binary, in which the run with neither rule gives 105; the exact run of the final
commit gives 104):
OBC_APPROACH_FLOOR).OBC_APRON).The list that stays
104 portals are in the gate of the spike.
logs/gate-list-r3.txthas each one with its parts (a portal can havetwo parts, so the rows sum to more).
Of the 40 low mouths, 22 are 90 % open and more, 12 are 75 to 90 % open, 6 are 50 to 75 % open and 0 are less.
30 of the 104 have one part only: a wall that stands there with no bores too. Without them the gate is 74 (7.3 % of
the entrances); without the 15 with no opening too it is 59 (5.8 %).
Notes on the parts:
more apart in height make a step on the edge of the one that wins. This step stands on
developtoday. The boresremove it at 149 portals (the cover and the lid take the place) and leave it at 36. It is storey 0's rule for
two ways and their heights (TrailView ground plan GP12.8: the heights contract for crossings, and the bake carries every shallow tunnel's tag #3349), not a rule of the bore.
crosses two entering ways (699), and another tube's cover on the entering way's edge (staggered twin tubes).
holds the least room over the bound's own height, and the floor's surface lies 5 to 35 cm over that height (270:
room 2.66 to 2.94 m of 3.0 m). The bound and the floor read the same line with two rules across and along it; the
build must give them one. Not traced further. The first spike held the cover at the bore's full height and had
slack here; 17 portals entered the gate with the least room.
least room over it, and the bake's levels say BORE (313, 195 to 201: a deck; 58, 59: a road on top that comes down
to the entering way). The body has no height there and lies under the lid; nothing stands out. These are faults of
the heights or of the levels. They go to TrailView ground plan GP12.8: the heights contract for crossings, and the bake carries every shallow tunnel's tag #3349 with this list. By the owner's rule for "no room" they are doors.
lies outside the corridor of the probe. Not looked at one by one.
The fallback
The gate does not come near 0 with the rules of this step. This is what to ship:
Of the 982 bodies, 72 stay in the gate: 30 with only a wall that
develophas too; 27 low mouths, of which 22 are90 % open and more, 4 are 75 to 90 % open and 1 is less; 15 with a wall that only the bores make or a step on the
opening. So 910 of 1,014 entrances pass the gate (89.7 %), and 940 (92.7 %) when storey 0's own walls do not count. The
target of 95 % is not met. The next gain is storey 0's rule for two ways, not a rule of the bore.
A. The joint
Unchanged from the first design, and confirmed by the review. Of the 1,033 joints of an entering way and a bore, 415
are cuts of one way and 618 are nodes of two parts; 42 turn by 10 degrees or more; 196 change width.
that edge once. No stage looks for the line afterwards.
entering way covers the patch on the other side. This reads labels of patches. It reads no distance.
edge at any heights. The floor keeps its own heights (round 1 gave the floor's corners the road's height; see
"What fixed which part").
radius. The frame's jamb (0.7 m) stands there; the corner is not rounded.
Options (a), (b) alone and (c) are rejected as in the first design ("Rejected options").
Decks (#3484) are not in this design. The review is right: rules 2 and 3 solve the turn and the T-junction
seams, not the rounded concave corner of the storey's outline, and not the green line in the viewer. A later deck
spike must test: the deck's storey in storey 0's slice, no round at a storey's open end, and the cost and the
digests in all 1,655 deck cells of box B.
B. The ground above
By rule 5 the roof is storey 0's own mesh, turned over. On each triangle the roof is linear, the lid is linear, and
at each corner the roof is at most the lid less the plate. So over the whole triangle the roof lies at least the
plate under the lid: no ground between floor and roof, and no roof over the ground, at the mesh's resolution. The
floor has the same caps, so the floor never stands over the ground and the roof never goes under the floor, at any
heights.
Only the corners of triangles over the same floor cap a roof corner. The road in front of a portal line, the
ground beside the floor and a face over another bore are no lid.
The cover's height. The cover has the ground's height (the slope rule), raised to the floor bound. The bound is
the bore's floor plus the least room, inside the bore's stroke up to its flat ends on the portal lines. It is not
capped by the terrain. So under the cover a mouth has the least room by construction, and more where the ground is
higher. The entering way's cut does not lower the roof: the bound lifts the cover over it, and the cover's own
corners carry the step (the end face over the opening).
The review's finding F3 names another cause: the cover's slope rule. That is not the cause. Where the bound holds a
point, the cover stands at the bound, whatever the cut does. The low mouths came from three faults of the bound (its
"near an end" test, its flat end, its cap at the terrain plus the lift) and from the caps ("What fixed which part").
With the bound off and the cover at the terrain or over it, as F3 proposes, the box has 386 low mouths
(section F).
What the lift costs. With no cap, the cover stands over the terrain where the bake says BORE and the device's
terrain is short. The measure is the run with the cap at
lift_max_m(on b954b62ac): the gate rises from 127 to191, with a low mouth at 64 more entrances; 7 of the 110 low mouths are under 50 % open.
logs/gate-list-r2.txtlists the 71 portals that are low only with the cap. These are places where the
bake's portal line lies in front of the first station that passes its test, or where the bake's ground is the body
of a way on top that the terrain does not hold. The bake rule that removes them: the portal line lies at the first
station that passes the BORE test, read on the device's terrain. Not built.
Cost per cell. No second slice, no second triangulation, no walk. Two triangles per storey-0 triangle over a
floor. A cell whose graph has no bore takes the path of a build with no bores (see "Evidence", cost).
C. The opening
One rule (rule 4): over a floor, storey 0 is a lid; every other face yields.
ground way is a lid of a lower way where, at each of its points whose stroke meets the lower way's stroke, its
surface lies at least that way's least room over that way's floor. A stretch is judged for each lower way apart.
Over a floor patch, a line's layers take part only where the stroke of its stretches on top of that floor's way
holds the patch.
bore's floor and cover hold the place. To give it to the entering way makes more walls than it removes ("What
fixed which part").
lose the place, and the place is a floor of the entering way's label under the lid. The opening is then the lid's
edge. This needs the paved layers of an entering way and of a way on top to name their way. The spike names them
along the whole line; the build names them on the stretch that is a lid or enters a bore, and
open_max_malongit. The reach must be checked with the deck comparison (section E).
The review's finding F6 is right: this is a sampled test on centre lines, and one failing point drops a stretch. It
reads only what the bake has: the solver's heights and the classes' widths. So the bake can decide it (section F).
D. The steps on the portal line
that yields. These are the wings of the end face.
the least room over or under the entering way is storey 0's own step ("The list that stays").
E. Decks away from tunnels
The cause of round 2's 8 cells stands as found:
crossingsran beforesplit, so a crossing read the level of thepart. Rule 1 removes it. The spike's "own" paved layers along a whole line make a second cause (12 cells of the
first spike differ far from every tunnel-tagged way). The build names them per stretch (section C). The deck
comparison "off | on" stays a proof count. It was not run again on this spike.
F. Size
Measured (gzip, wasm and glue,
wasm-opt -O3, the guard's number):trim-wt), offThe trimmed spike is a spike with its dead paths removed. It is not a written build: its "off" is 6.3 kB over round
2's "off" (code with no feature gate), and its raw difference of 54 kB holds the iterator adapters, the map drop
glue and the functions that the feature moves out of line, as the review lists for the first spike (about 20 kB raw).
Estimate for a written build of the body without the frame. The basis is the measured 23.7 kB, less the review's
list of what a trim removes (about 20 kB raw, 9 kB gzip):
The three cuts of the review, each with its census run over all 1,014 portals:
OBC_NO_BOUND=1: gate 452. 338 entrances enter the gate, each with a low mouth; 386 low mouths in all, 33 of them under 50 % open and 70 between 50 and 75 %. With the bound kept and capped atlift_max_m(OBC_LIFT=clamp): gate 191, 64 more low mouths.OBC_NO_WEDGE=1with today's lines: gate 160. 42 entrances enter the gate with a wall across the way (33 of class Wa); wall within 3 m at 226 portals (173). The bake's move is not built, so its look has no run.Recommended: cuts 1 and 3 in the bake, cut 2 not. The body is then about 11 kB (10 to 13). The target of 10 kB
is at the edge of the estimate. What the kilobytes over it buy: the floor bound (1.4 kB) keeps 338
entrances open. With the frame (3 to 4 kB) the chain of bridges and tunnels stands at about 64 kB against about
60 kB.
G. The simplest alternative
The strongest simpler case is the review's: the bake decides, the device draws. The bake decides the portal
lines, the lids and the body's length, with 0 device bytes and one decision for all cells. The device keeps rules 2,
3 and 5 and the floor bound. This design takes that case for the lids and for the line in front of a lid (section F,
cuts 1 and 3). It does not take "no floor bound": the bound is what makes a mouth open by construction, and the
census shows its price.
A dark quad on the portal line with no body is rejected as before: 345 of the 1,014 portals are underpasses, and an
underpass is seen through.
H. The build plan
Each slice leaves a valid state. Nothing draws before step 3.
lids: the metres along and the lowerfeature of each stretch on top); a portal line in front of every lid that reaches over the entering way; the
portal line at the first station that passes the BORE test.
levelsonly; a new features revision; the fixtureand box B rebaked by the step's own script. The closed lines of rule 6 are no doors (a BORE | HIDDEN line inside a
bore is a plain end wall), and they stay out of the list of HIDDEN runs of step 1. Proof: the tags of tunnel ways
differ only by the new lines and the lids; the levels test of the fixture. No pixel change.
bores: the foundation and the body (rules 1 to 5), from the branchclaude/gp12-4-bore-storeywith the spike as the model, written new and small. No pixel change with the feature off (256 digests).
bores: the frame and the end face (step 2b): a prism on the opening's pieces.bound; a corner of the storey mesh is the least of its label's height and its caps, floor corners too; a bore's
mesh is built when the cell's mesh is built, because it is made of it. The bridge keeps the triangles over a
floor, not the cell's mesh, and nothing in a cell with no bore.
What stays of the branch. The graph (the tunnel tag on ground kinds,
splitby states, the junction rule,levelled), features revision 22 and the rebaked fixture, the storey of a bore way and its bands,DeckWayfor abore, the signed storey in the records' merge, the caps of a corner and
corner_height, the surfaceCover, the edgekind
Portal, the floor bound's walk, the changes oftests/graph.rs.What goes.
PortalMap(walk, back) and the floor's relabel map;inner_grid; the bore path ofmesh_storeys(its own triangulation); the storey's own slice for bores; the cover's band candidate; the test "in front of the
line" in
decks; the tolerance test of a portal edge; the cap of the bound at the terrain plus the lift; the sill'srelabel of the floor's corners.
The proof counts, by
probe/census4.rsover all 1,014 portals:paved_overlists each wayAlso: 256 digests with the feature off; the digests of storey 0 in the cells with no bore, on against off; the deck
comparison "off | on" over box B; the WASM size "off | on"; the cell times "off | on" on an idle machine.
The pictures of the build: the 12 places of "What a rider sees", and the 10 worst portals of the build's own
gate list. A debug oblique cannot prove an open mouth; the eye-height renders of step 3 must.
What a rider sees
12 debug obliques of the spike 8938a7348, each opened by the designer (
pics/final/INDEX.md; flat colours,orthographic, no viewer render). Seven places are in the gate (the worst of its parts); five pass it: four are the
review's worst places, and 880 left the gate with the last commit.
Of the 12: read right or open 7, not clear 2, closed or broken 3.
Invariants
triangle. The floor lies under the lid too.
bore's class. (The spike misses it by 5 to 35 cm at about 22 portals; see "The list that stays".)
faces and mesh are byte-equal with the feature on.
true in the spike (the "own" layers per line).
The review's findings, checked
hdbg: caps of label 10 over the railway's floor); rule 4.decksgave 4 segments 27 m out withnear [true, true]; after the fix the mouth is full. F3's own proposal as a census run: section F, cut 2.Rejected options
lift_max_mon the deviceLimits and risks
device with the current lines. A lid tag keeps the look by construction (the same test on the same samples). A
line in front of a lid changes the look: the bore's box stands square in front of an oblique way on top, under
cover.
over the floor. The clamp run lists the places.
lower rank is cut where they overlap. With rule 6 this is rare.
cut at an existing end sample makes no new piece).
line; it keeps a second merge behind a switch.
Evidence
All on timosPC, test box B (
tunnels/step2/bake/fxb; 1,014 census portals). Round 2 isclaude/gp12-4-bore-storey4a42b8f. The first spike is 04243fc8a (the behaviour of 86061544e). This spike is 8938a7348 on the local branch
claude/gp12-4-joint-spikeintunnels/step2-redesign/wt. It is b954b62ac with the two approach rules off; theruns of the cuts, the 30 m body, the size and the digests are of b954b62ac (with the first approach rule on).
trim-wtholds b954b62ac less its test switches (local branchclaude/gp12-4-joint-spike-trim76eba61e2). Files are in/media/timo/OSM1/tmp/ground-plan/tunnels/step2-redesign/;STATE.mdthere has the run log.probe/census4.rs(census4r2.rsfor round 2's API), runnerscripts/c4.sh, tablesscripts/gt.pyandscripts/stay.py. Runs:logs/c4-box-off.json(bores off, the spike's pieces),logs/c4-box-round2.json,logs/c4-box-r1.json,logs/c4-box-r3.json(this spike, 8938a7348; the same file asc4-box-final.json),logs/c4-box-r2.json(b954b62ac), and of b954b62ac with a switch:logs/c4-box-r2-body30.json(OBC_BODY_M=30),logs/c4-box-r2-clamp.json(OBC_LIFT=clamp),logs/c4-box-r2-nobound.json(OBC_NO_BOUND=1),logs/c4-box-r2-nowedge.json(OBC_NO_WEDGE=1);logs/c4-box-r3-apron.json(the second approach rule on) andlogs/c4-box-r3-switch.json(neither rule, thesame binary). Round 2's pieces have no fixed side; the probe turns them by the
census portal's direction.
probe/hdbg.rs(the height rule's parts at a point, the mesh corners and caps about it) andprobe/spikedbg.rs(the faces and portal edges at a point); try runslogs/c4-try-t1.jsonto-t10.json,logs/c4-try-noapproach.json.misses are 0 at every portal of the spike. Round 2: ground between floor and roof at 177 portals, roof over the ground at 184, planar seams at 314, misses at 542, closure 0.
logs/digests-off-r2.txtagainstlogs/digests-off-spike.txt).faces, far block, rim and mesh are byte-equal on against off (
logs/digests-on-r2.txt). In the 25 of them with adeck, the deck mesh's bytes differ and its triangle count, ways and ends do not: its layout holds the caps' count
with the bores. The code asks the graph for a bore first (
faces::any_bore), so these cells take the same path.the 37 cells of the 40 subset portals (95th 2.25 times), with full-length bodies. This spike adds the lid strokes
per line and one more cap per floor corner.
pkg/off-r2,pkg/on-r2,pkg/off-trim,pkg/on-trim; logslogs/wasm-*.log.pics/final/INDEX.md.DESIGN-r1.md, Evidence.Not measured
The bake's rule 6, the lid tags and the moved portal lines (bake code). The deck comparison and the crate tests on
this spike. The cell times on this spike. The count for invariant 7. Any viewer render. Each portal of the parts
"wall only with the bores", "step", "closed" and "none".
Generated by Claude Code
timohueser commentedon Nov 20, 2026
Closed in the backlog cleanup as superseded by #3651. That epic rejects the method of this issue by name ("continue to derive the shapes on the device (#3345, draft #3643)"), and P8 (#3685) deletes the
boresfeature that the merged pieces (#3488, #3531, #3566, #3572) live behind. Its useful parts (the count of test entrances, the rule that a way on top keeps its paved layers, the light) carry over into #3651.Generated by Claude Code