Skip to content

TrailView ground plan GP12.2b: the ground at bridge ends: the cause of today's torn walls, ordinary ends, end faces, the ceiling under a deck #3352

Closed

Description

@timohueser

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, the gallery, the costs, the owner's decisions) and part 1's proof on timosPC:
/media/timo/OSM1/tmp/ground-plan/crossings/gp122b/PROOF.md (summary in the comment on this issue). Short names:
gp/ = host/obc-ground-plan/src/, tv/ = builder/web/src/lib/trailview/. Code places verified on ground-plan
8f03b58.

Goal

A way that runs onto a bridge ends as a way ends on the ground: its paved outline ends flat on its end line, its bank
runs round the end and down under the deck, and on the end line stands a plain concrete end face, at least as tall as
the deck plus a gap. Under a deck the ground stays under the deck, so an opening shows where the heights give room. A
step that another way's cut holds at a way's edge is at most a wall high; a taller one is counted as a fault.

Part 1 (done): what the proof found

  • The no-room rule fit is sound. No torn or tall wall at W1, T01, R02, W5, R11 comes from it.
  • fit with the structure height as the end edge's cap does not make an end face: at every deck end of the fixture
    (34 of 34) and of box B (82 of 82) it makes a bank that stands 0.7 to 1.35 m over the deck's underside on the end
    line. The earlier part 2 failed its own invariant at every deck end.
  • The other faults: a Held step has no size limit (W1 6.6 m, T01 round 1 7.3 m at a rail's edge); at W1 the
    deck end lies on the lower way (0 to 4 m from the cycleway, 7.3 m over it); the stamp's tunnel COVER lifts the
    ground at R02, W5, R11; the 0.5 m stone rule of tv/plan/piece.wgsl:163 draws a 0.6 m step as stone (T01).
  • Part 2 below carries the four changes that the proof asks for. Change 3 (COVER) belongs to GP12.4.

Depends on / blocks

  • Depends on: GP12.2a (merged). It does not read GP12.1's levels: a deck end is a level-change end whose way beyond
    is a deck way (Level::Deck), at today's tag end.
  • Blocks: GP12.3a (its storey records start from the ceiling lines of this step), GP12.3b (it takes rule 2b: the end
    line clears the lower way's reach), GP12.4 (it reuses the end face and the ceiling in its bore form, and stops the
    tunnel COVER).

Part 2: the build (plan path only)

Everything below runs only with ground=plan. The default path does not change until GP19 (#3110).

Change 1. The ceiling bounds the ground: the end face exists by construction

What exists. A line's end at a level change has the end kind LEVEL (line words 12 and 13,
gp/faces/records/mod.rs:146-150, set at about :830-840). Its strokes end square there, the slope layer too
(ends in candidates, gp/faces/mod.rs:1079-1102). The slope band fades into the terrain over band_fade_m
(gp/faces/heights/mod.rs:1322), and step() labels the step there StructureEnd (:1491-1501). The near records
hold the lines of storey 0 only (in_level, gp/faces/mod.rs:197-206): nothing in them knows a deck. On the plan
path the corners clamp the ground to the old stamp (stamped_ground, stamp_bounded,
tv/terrain/stamp-core.wgsl:332-338, read by height_of in tv/plan/corners.wgsl:64-80), whose deck terms hold the
ground under a deck today.

What is new.

  1. Deck end kind. A LEVEL end whose way beyond is a deck way takes the new end kind DECK (3) in the same words.
    Tunnel and gallery ends keep LEVEL and today's end until GP12.4 and GP12.7.
  2. Round slope cap. At a DECK end the paved layers (carriageway, shoulder, verge) keep their flat cap; the slope
    layer takes the round cap that a free end has. So the bank runs round the end. The band fade at a deck end goes.
  3. Ceiling lines in the near records. Each way of Level::Deck whose stroke plus ceiling_release_m meets the
    cell's window is a line of the lines table with the new flag CEILING (finish word bit 12, free today) and its
    drop in word 10 (the return grade, which a ceiling line does not use): drop = deck depth + ceiling_gap_m. The
    deck depth is the tile's deck tag, else the class's deck_least_m (the rule of deck_depth in
    gp/graph/levels.rs:570; move that function so that the records and the levels share it). The bins list a ceiling
    line's segments within ceiling_release_m + HEIGHT_PAD_M of their square, with the height segments. Every reader
    of storey 0 (candidates, paved field, slopes, outline edges, junctions) skips a CEILING line. MAGIC changes.
  4. The bound. For every storey-0 point of the rule Terrain or Slope (the corners and the pages): ground =
    min(ground, ceiling). The ceiling at q is, over the ceiling lines in q's bin: the line's surface at q's foot less its
    drop, inside the deck's stroke (paved half-width + edge_band_m, flat at its ends); beside the stroke it rises at
    face_max per metre; it lets go smoothly and is gone at ceiling_release_m, so it never steps. The slope rule
    reads the plain terrain as today; only its result is bounded. A paved face and a water face keep their heights.
    Rust: height_at and by_label (gp/faces/heights/mod.rs:1372, :1626); WGSL: the port in
    tv/terrain/heights.wgsl (height_at, :1076), which the corners and the pages (tv/terrain/model.wgsl) share.
  5. The end face. On a DECK end line, the ground just past the line is at most the underside less the gap, so the
    step on the line is at least deck depth + ceiling_gap_m; fit adds to it where the need is larger (the end
    edges' cap stays Line::wall_max, :379-385). step() labels EndFace (new) every step on a DECK end line,
    whatever its lower side (at W1 it is the cycleway's paved face), and every step whose lower side the ceiling
    holds (the Cause of a side says when the ceiling bound it). StructureEnd stays for tunnel and gallery
    ends only.
  6. The old deck terms stop. nodeStamps (tv/terrain/bins.ts:1048) writes no record for a BRIDGE line with
    plan on the levels that the records cover (node.level >= NEAR, 16): the stamp's deck terms (DECK, TROUGH,
    ABUTMENT_CELLS, CLEAR_M, TROUGH_M, TROUGH_REACH_M, DECK_GAP_M:
    tv/terrain/stamp-core.wgsl:87, 101, 150-159, 671-685) get no input there. The far levels keep the trough until
    GP12.6. bridges.ts builds no abutment box on the plan path (abutment(), tv/structures/bridges.ts:119). The old
    deck itself still draws until GP12.3b.

Data. Rule values in plan.json heights (so HeightRules carries them in the block's rule words):
ceiling_gap_m 0.5 (#3105 invariant 1), ceiling_release_m 8. Bytes: per deck way, one line (64 B), its samples
(48 B each) and bin items; estimate 0.6 kB per 20 m deck, 15 kB for a viaduct of 500 m. The outline edge record does
not change.

Deleted from the earlier design. The end edges' cap set to the structure height (part 1: it gives a bank, not an
end face); the ceiling on the terrain that fit reads (T' = min(T, ceiling)); the storey field (GP12.3a adds
storeys; part 2 needs only "not storey 0"); ceiling_gap_water_m (see Later).

Change 2. A Held step gets the wall's size limit; the deck end must clear the lower way's reach

What exists. In slope() (gp/faces/heights/mod.rs:1254-1330) a line's cut rises toward its budget (the soft
top of the fills that counts admits) up to steep: its wall (a wall class, up to wall_max_m) and past the wall's
top its steepest cut, at most face_max (groups, :1217-1224). Where the gap to the higher line is too short, the
rest stands at the higher line's edge. step() labels it Held with no size limit (:1471-1481). counts
(:1236-1238) leaves a rail out of every budget, so a rail's edge holds a step of any height (T01 round 1: 7.3 m).

What is new.

  • 2a. Size limit. step() labels a step Held only up to the held line's cap (Line::wall_max: wall_max_m for
    a wall class, 0 for another) + step_m. A larger step is a fault that the new check held_over_cap
    (gp/faces/heights/check.rs) lists with its place, its height and the gap between the two paved outlines. The face
    above the cap is today's steep: a wall up to the cap, then the steepest cut. This step adds no steeper face.
  • 2b-rail. A rail counts with its fill less its cap. In the budget of another line, a rail's fill counts less the
    rail's cap (wall_max_m). A rail bed up to 4.5 m high keeps its own step as today (GP3's rule: a road gets no wall
    from a low rail bed); a higher one lifts the road's cut by the rest: the road's wall up to its cap, the steep face
    above it. Rust (counts, the soft top in slope) and WGSL (h_slope, tv/terrain/heights.wgsl:890) line for
    line. No data.
  • 2b. The deck end clears the lower way's reach (rule for GP12.3b). No step of storey 0 can carry W1: the ramp's
    paved end touches the cycleway's paved face, 7.3 m over it, so no bank, wall or face fits (proof: with the
    cycleway's cone left out the step moves to the cycleway's edge). The end must move: GP12.3b places A where the end's
    bank toe (the end's height over the lower way, at the class fill) lies outside the lower way's paved stroke, not
    only where the deck stroke clears the ground arms; the stretch between the tag end and A draws as deck, on the
    way's own heights. This step only counts such ends (proof: 6 on the fixture, 20 on box B) and lists them.

Data. None. Deleted. Nothing.

Change 3. The stamp's tunnel COVER is GP12.4's

The stamp's tunnel term COVER (tv/terrain/stamp-core.wgsl:102, 164, 690) lifts the ground beside the roads at
R02 and W5 (+10.2 m and +4.8 m against the crate's heights) and adds height at R11 round 1. This step does not
touch it, and judges no end face at those places. GP12.4 (#3345, its section 4) stops it on the plan path.

Change 4. The end face's look keys on the edge kind

What exists. A step slot is stone (MAT_WALL) when a corner lies on a wall class's road and the step is 0.5 m or
more (WALL_STEP_M, tv/plan/piece.wgsl:27, 163); else a bank. A slot carries no edge kind: its look comes from its
labels and its height (the mesh's module comment, gp/mesh/mod.rs:43-50). The face edges have a kind
(EdgeKind, gp/faces/types.rs:46-54), set only to Continuous or CellBorder (gp/faces/mod.rs:1741).
DeckEdge exists and is unused.

What is new. The faces give an edge on a DECK end line the kind DeckEdge. The mesh puts the slots on
DeckEdge edges last, in runs of their own, and the head counts them (one more HEAD word: 4 B per cell; the mesh's
MAGIC changes). slot_vertex draws a slot of those runs as an end face: plain concrete, no pattern repeat, at any
height, never stone. All other slots keep today's rule.

Later (cut as polish under the owner's guidance; each is safe to leave)

What Why it is safe to leave Owner of it
The need C1 along the rim (#3379 open point 1) It makes faint shading bands on cut slopes (m2-eye, v04), not steps; the gates hold known-faults list
A piece's host change snaps to a bin cut (#3379 open point 2) It moves a cone with the bin grid (half-bin shift 0.289 m); the border check holds, so two cells agree known-faults list
A wall-top profile beyond #3379 An end face's top is the flat paved end; nothing steps there none needed
A coping band on the end face; railings on an end face under a way A plain face is whole and closed; ornament GP12.5, #3115
The gap of 0.3 m over water A water face keeps its own height; the bank under a low deck lies 0.5 m under it. The step counts decks over water with less room than depth + gap #3107, #3108
The stone look keyed on the step kind for every step Needs a kind per slot; today only the end face changes. T01's 0.6 m wedge is a heights fault (two carriageways 0.6 m apart) GP12.8
A face steeper than face_max where two ways touch No face fits at a gap of 0; at deck ends rule 2b removes it GP12.3b; known-faults list
The short side steps where the ceiling rises beside the deck's corners (under deck + gap, under 0.6 m long) They keep today's look; they close by construction none needed
A paved face of storey 0 that stands over a ceiling (no room) The ladder makes those ways ground ways GP12.3b

Invariants

Each one has its check; the crate tests run over every cell of the Engelberg fixture, and over the box-B places'
fixture of part 1 (gp122b/fx-today) where named.

  1. No bank over the underside at an end line. On every DECK end line (34 on the fixture), no storey-0 point of
    rule Terrain or Slope within 0.5 m past the line stands over the underside less ceiling_gap_m. Check: crate
    test, the numeric test of part 1 made a test, over the fixture and fx-today (82 deck ends); target 0 of all.
  2. The end face. The step on every DECK end line is at least deck depth + ceiling_gap_m − step_m along the
    whole paved end, and its label is EndFace. No StructureEnd at a deck end. Check: the same test.
  3. The ground under a deck. Inside every deck's stroke, every storey-0 point of rule Terrain or Slope lies at
    least ceiling_gap_m under the underside. Paved and water faces are not checked (they keep their heights); their
    count over the ceiling is reported. Check: crate test on a 0.5 m grid over the deck cells.
  4. No Held step over the cap with room for the steep face. Every step that held_over_cap lists has a gap
    between the two paved outlines under (its height − both caps) / face_max. Check: crate test over the fixture;
    the count is reported and gated at the measured value; the list goes into the pull request.
  5. Every ending is whole (TrailView ground plan GP12: bridges, tunnels, underpasses and galleries: decks, bores and roofs with the ladder (the contract of the crossing block) #3105 invariant 3). Ground triangles within 3 m of a DECK end line that are steeper
    than face_max where the terrain is not, and are not step slots: 0. Check: crate test over the fixture.
  6. No new torn face. The gates of host/obc-ground-plan/tests/heights_fixture.rs hold: steps with no rule
    (EDGE_MAX), steep pairs (STEEP_MAX), none over 6 m/m, wall_steps, slope_steps, inside. A gate that moves
    is stated in the pull request with its cause.
  7. Cell independence. The ceiling reads only the deck line's own samples through the bins, as the paved field
    does. Check: the border check of heights_fixture.rs (shared border points: 0 differ) with the ceiling lines in.
  8. The end face looks like one. Every slot on a DeckEdge edge is in the end-face runs; none draws stone. Check:
    crate test on the mesh of the fixture's deck cells.
  9. Rust = WGSL. Check: the model probe tests/browser/trailview/model-probe.mjs at the fixture's bridges.
  10. The default path does not change. Check: the default-path views give the same bytes.

Pictures required before review

  • The crossing gallery, all 50 places (crossings/gallery/shoot.sh), on the demo data that the owner flies at the
    time of the shoot. One sheet per place: today | after, today = ground-plan at the branch point, after = the
    pull request head, the same data in both columns; rows eye and obl; each tile with its label box (build, data,
    place, camera). The implementer opens every sheet and writes a verdict per place beside gallery/INDEX.md.
    • Must improve: S5 (an opening under the railway deck, no grey wall), B01 and p08 (no grey band under the
      decks; their far views: nothing worse), W2 (no pale piece at the deck end), p02 (no abutment block).
    • W1: the end faces are whole concrete faces; the shards beside the ramp ends stay until GP12.3b (rule 2b). It
      must not get worse.
    • Unchanged, judged after GP12.4: R02, W5, R11, T01 (T01 round 1: the rail step may shrink by 2b-rail).
    • Must stay GOOD: B06, B07, B12, B14, N3, B1. A top view at the deck ends of the five GOOD deck places:
      no slit between the end face and the old deck.
  • The debug tool's top-down pictures (layers faces,heights,mesh) at bridge-top (46.821463, 8.376719),
    under-bridge-top (46.815033, 8.394683), stream-under-bridge-top (46.795161, 8.385839).
  • The owner sees the sheets and gives his OK before the merge.

Cost to measure

Base: ground-plan at the branch point; develop's head for the frame. One quiet run each.

  • Cell build time native, p50 and p95, over the fixture and over its deck cells (examples/debug --bench).
  • The near block's bytes per cell (sum, p50, p95, max) and the mesh head's word.
  • The page time and the corners' pass on the GPU (plan-timing.html): the pages read the ceilings per texel.
  • Frame: walk, valley-drive (flight.mjs --gpu, clock held, GPU lock) and luz (47.05135, 8.30525, heading 60,
    eye 1.7).
  • WASM bytes, gzip and raw: the step's share of the block's 15 to 25 kB.

Stop rule

Stop and report to the orchestrator, with pictures and numbers, if:

  • the ceiling moves a paved or water face, or an end face needs a repair pass or a shape of its own;
  • invariant 1, 2 or 3 fails at a deck end and no general rule explains it;
  • a GOOD place gets worse, or a heights gate grows, and no general rule explains it;
  • a cost reaches twice its base, or the block's WASM passes 25 kB;
  • two review rounds fail.

What this step must not do

  • No tunnel or gallery work: their ends keep LEVEL, the square end, the band fade and StructureEnd; no COVER,
    portal, hole or bore change (GP12.4, GP12.7).
  • No deck sheets, storey faces, shell, parapets or piers (GP12.3a, GP12.3b). The old deck draws as today.
  • No end line other than today's tag end; no call of graph::levels from the device build.
  • No change to the ways' heights, to the solver, or to ground plan: the slope and the wall are values of the rim's vertices (GP12.2a) #3379's rim values; no steeper face than steep.
  • No new browser test; no change to the default path.

Decisions already made by the owner that apply

  • "End walls and tunnel frames taller than 4.5 m: yes of course."
  • Walls only on road and rail classes, at most 4.5 m, then a steep face; a path gets banks.
  • Torn or triangular faces are the worst fault; every ending gets a proper end piece. No special cases, no repair
    passes.
  • Finish the epic; about 95 % good is enough; record the remaining faults on the known-faults list instead of a new
    round.

Requirements: the verification console could not be checked from this PC (no agent token here). The orchestrator sets
the pull request's Requirements: line.

Activity

  1. timohueser commented on Nov 11, 2026

    @timohueser
    OwnerAuthor

    From GP12.1 (PR #3367) after its review

    GP12.2b does not read GP12.1. For the record: the end-line merging of a sheet and the sheets/storeys left GP12.1 (they wait for the step whose pictures decide them); the code is on branch claude/gp12-1-later (68b3ddd). graph::build does not call graph::levels yet; GP12.3a links it in (target at most 12 kB of WASM gzip).


    Generated by Claude Code

  2. timohueser commented on Nov 11, 2026

    @timohueser
    OwnerAuthor

    GP12.2a (#3343, PR #3379) changes what the end faces of GP12.2b take: an outline edge has 9 words now (the grade at b and the flags ON_A/ON_B in word 8, MAGIC 0x4f425213). The slope and the wall are values of the rim's vertices; an end where the straight rim goes on (a cut or a split) has no round cone, a real corner keeps it. A level-change end stays a corner. Details: #3343


    Generated by Claude Code

  3. timohueser commented on Nov 12, 2026

    @timohueser
    OwnerAuthor

    GP12.2a (#3343, PR #3379) is ready for the owner's decision. Its final record and two open points go to this step.

    The record that the end faces take. An outline edge has 10 words (MAGIC 0x4f425214). Word 8 holds the grade at b (low half) and the turn at b (high half). Word 9 holds the turn at a (low half). A turn is an i16 in steps of π/32767 to the one edge of the same host that goes on from the end with the same paved height bits; i16::MIN means that no edge goes on. The flags ON_A/ON_B of the earlier comment are gone. The slope and the wall are linear along a piece between the vertex values. Past an end's normal the line falls at a rock face's slope. The end's own cone covers only the fan of a turn toward the paved face, and all past a free end. A level-change end with no edge going on is a free end.

    Open point 1: the need is C0, not C1 along the rim. The need is linear per piece, so its gradient along the rim jumps at each vertex. Where the need changes fast (a rail cut), the ground's grade along the rim jumps by about Δ(dneed/da) · off / run at the vertex's normal. This is the likely cause of the two views of #3379 that did not improve: m2-eye (a dark band with a straight vertical edge on the rail cut) and v04-wall-road-junction (a sharper crease with a small hook). The first review proposes a general fix: split a piece while the need at its middle leaves the line between its ends by more than a few centimetres. That is cheaper than a C1 interpolation, which must read the neighbours.

    Open point 2: a piece's host change snaps to a bin cut. A bin piece takes its host from the paved owner at its middle (records/mod.rs, outline_edges). So a change of host along a rim moves to the nearest bin cut, and both ends there get free-end full cones. Both move with the bin grid. This is one candidate for the remaining half-bin shift of #3379 (worst 0.289 m against the target 0.0 m).

    Sources on timosPC: crossings/REVIEW-3379.md (N4), crossings/REVIEW-3379-delta.md (N7), the sheets crossings/gp122a-pics/sheets2/.


    Generated by Claude Code

  4. timohueser commented on Nov 12, 2026

    @timohueser
    OwnerAuthor

    GP12.2b part 1: the proof (on #3379 head b354706)

    Full report on timosPC: crossings/gp122b/PROOF.md (sheets gp122b/sheets/, sections gp122b/sections/, probes gp122b/scripts/).

    Method. A crate fixture of the five places from the demo release and from round 1; the heights scan within 70 m of each place (every step of 2 m or more, its step() label, the outline edges with h, t, need and fit). Two switch probes: a dist in which the plan's height_of ignores the old stamp (tv/plan/corners.wgsl:78-79), and the lower way's cone left out. The GPU drawn terrain (trailView.check) on the same sections.

    The five places

    Place Visible wall Cause and proving number Kind
    W1 shards at the link ramps' ends The ramps' ground starts at the bridge tag end, 0 to 4 m from the cycleway and 7.3 m above it. The cycleway's cut holds the ground: a Held step of 6.57 m (47.0379687, 8.2962628). With the cycleway's cone left out, the step moves to the cycleway's edge (7.08 m). Held has no size limit (gp/faces/heights/mod.rs:1475-1481). the plan's slope rule, set off by the place of the deck end
    T01 stone wedge on the carriageway Unrelated step of 0.45 to 0.65 m between two motorway lines; stone only by the 0.5 m rule of tv/plan/piece.wgsl:163 heights + look rule
    T01 stone at the left jamb NoRoom cut walls of 4.5 m; gone with #3379 the rule (fixed)
    T01 (hidden in the frame) StructureEnd step of 12.1 m at the tunnel tag end (band fade, :1322); shows when the stamp is off the plan's square end
    T01 r1 7.3 m wall between the narrow gauge and a secondary road Held: a rail does not count in a road's budget (counts, :1236), no cap the plan's slope rule
    R02 stone wall taller than a house The stamp's tunnel COVER (stamp-core.wgsl:686-691) lifts the slope beside the road. Crate: no step of 2 m or more within 70 m; on one section the crate gives 443.3 m, the drawn terrain 453.4 m (+10.2 m). Probe without the stamp: wall gone (29 % of the eye picture changes, both data sets) old structure code
    W5 trench between stone walls the same COVER: crate 446.7 m, drawn 451.5 m (+4.8 m); gone in the probe (16 to 19 %) old structure code
    R11 grey towers; stone wall at the lower left the old portal box (probe: no change); the tertiary road's own cut wall at the cap (need 5.5 m, wall 4.51 m) old structure code; the rule, right
    R11 r1 stone face taller than a house #3379 turns it into a capped wall + ground face; a 4 m stone triangle stays (Held 4.22 m; the stamp makes it taller) the plan's slope rule + stamp

    Six distinct causes; none is fit itself. #3379 at these places: better at T01 and R11, the same at W1, R02, W5 (changed pixels 0.4 to 2.2 %, R11 round 1 12.9 %). On an end line #3379 reads t beside the way's end (the frame direction from the end sample is across), not under the deck.

    The end-face rule on every deck end

    fit with the end edge's cap set to the structure height, with t under a ceiling (deck underside − 0.5 m, least deck depth), with GP12.1's levels:

    fixture box B demo box B round 1
    bridge-way ends (DECK state) 109 (34) 285 (82) 286 (85)
    bank / wall ≤ 4.5 m / wall > 4.5 m 108 / 1 / 0 283 / 2 / 0 284 / 2 / 0
    DECK ends whose bank stands over the deck's underside 34 of 34 82 of 82 84 of 85
    ... over (median / max) 0.7 / 1.35 m 0.7 / 1.35 m 0.7 / 1.35 m
    DECK ends whose bank toe reaches the lower way 6 20 19
    wall top along the end line > 1 m/m 0 0 1

    An end face appears only where the need passes fill_max × run (10.1 m road, 5.55 m path); the largest need at a DECK end is 8.8 m. No negative or huge wall. Worst ten in PROOF.md.

    Conclusion

    The no-room rule is sound, but as part 2 specifies it, it does not make end faces: it makes banks that run into the deck at every deck end (invariant 3 fails everywhere). Yes, with these changes first:

    1. The ceiling bounds the ground itself (ground = min(ground, ceiling) inside the deck's outline, rising beside it), not only the terrain that the rule reads. The end face then exists by construction (at least deck depth + gap), and fit adds to it.
    2. A Held step gets the wall rule's limit with the steep face above it; the deck end clears the lower way's reach, not only its stroke (GP12.3b's line A).
    3. R02 and W5 are the stamp's tunnel term COVER, which part 2 item 4 does not turn off: judge them after GP12.4.
    4. The end-face look keys on the edge kind, not on the 0.5 m step rule.

    For tunnel mouths (#3345): the open cut can use the rule as it is (R11: a right 4.5 m cut wall); the portal needs change 1 in its bore form and must replace the square end with its band fade (the 12 m face at T01), with COVER off first.

    Stop rule: part 1 shows that the end-face rule as written leaves the ground over the deck at every end; this goes to the orchestrator before part 2.


    Generated by Claude Code

  5. timohueser commented on Nov 12, 2026

    @timohueser
    OwnerAuthor

    GP12.2b part 2: design note (before the build)

    Branch claude/gp12-2b-deck-ends from ground-plan 8f03b58. Short names as in the body.

    1. Deck end kind. gp/faces/mod.rs lines() marks an end deck where the end station is a joint or junction with an arm of Level::Deck (the arm levels of the graph; no graph::levels). records::line_words writes DECK = 3 in line words 12 and 13 where it wrote LEVEL and the flag is set. Words 12 and 13 hold 0 to 2 today, so 3 is free.
    2. Round slope cap. candidates(): for the slope layer only, ends adds the deck ends; the paved layers keep cut(). The square slope end goes, so the band fade at the end line goes with it (no special code).
    3. Ceiling lines. New faces::ceiling_lines = lines() with the filter Level::Deck (one leg). records::encode appends them after every storey-0 line: finish bit 12 CEILING (free: the fields are 0-7, 8-11, 14, 16-23, 24-31), word 10 = drop = deck depth + ceiling_gap_m, sample half-width = half + widen + edge_band_m (the stroke, so no new word). Their height segments go last in each bin. deck_depth moves to WayGraph::deck_depth (gp/graph/mod.rs); levels.rs calls it. Records MAGIC 0x4f425215.
    4. The bound. heights::ceiling(b, q, segs) walks the bin's height segments from the end while the line is CEILING: surface at the foot less the drop, + face_max × metres outside the stroke (flat at the line's ends), + the cone release term, gone at ceiling_release_m. by_label takes min for rule 0 and 1 and sets Cause::ceiling. present, nearest_way (tv/plan/corners.wgsl:91) and the outline ends skip CEILING lines. WGSL: h_ceiling in tv/terrain/heights.wgsl, used by height_relative for M_NONE, rule 0 and rule 1.
    5. Rules. ceiling_gap_m 0.5 and ceiling_release_m 8 append to RULE_FIELDS (block words HEAD+21, +22; the legs word moves to HEAD+23).
    6. Labels. step() first: EndFace on a DECK end line (|past end| <= END_UNITS) or where the lower side's cause is the ceiling. A Held step over the paved side's host cap (+ climb + step_m) gets the new label HeldOverCap; check::held_over_cap lists those (place, height, gap to the holding line's outline). 2b-rail: in slope (Rust) and h_slope (WGSL) a rail counts in another line's budget with its fill less its wall_max.
    7. Stamp and abutment. nodeStamps treats a BRIDGE line with plan at level >= NEAR as a records way (no record; it still marks heights and the mask). structureMeshes gets plan from the worker message (already sent) and bridges() skips abutment() with it.

    What differs from the issue (verified in the code):

    • The mesh reads no EdgeKind today (confirmed): mesh/points.rs input() drops e.kind, and cdt::Segment has no kind. Plumbing: points::Input gets one flag per face-edge piece (the edge's kind is DeckEdge); mesh_cell puts those slots after the others in runs of their own and counts them in head word 10 (HEAD 11, MAGIC 0x4f424d04). The faces set DeckEdge in Sliced::cell (both builds pass there, so the records test's cell-only check still holds) for an edge whose ends lie on a DECK end line. Viewer: the Piece uniform word steps (no shader reads it today) becomes the first end-face slot; slot_vertex draws a slot from there as MAT_CAP (the plain concrete of a wall's cap), never MAT_WALL. No new pipeline or draw call.
    • The paved end is carriageway + shoulders (rule Owner). The verge is a Slope face (BAND_LABELS), but every class has verge_m 0 today.
    • height_at returns the plain terrain for rule 0 without a bin read (gp/faces/heights/mod.rs:1374, WGSL height_relative): the bound adds that read. A terrain point without a ceiling in its bin reads the bin's last height segment and stops.
    • Terrain-rule corners read the drawn page (tv/plan/corners.wgsl:66), not height_at: the page pass (model_height) bounds them.
    • rules.rs checks region_fade_m as the last rule field: it reads it by name now.

    Generated by Claude Code

  6. timohueser commented on Nov 12, 2026

    @timohueser
    OwnerAuthor

    GP12.2b part 2: deviations in PR #3396

    1. Stroke width. The ceiling's stroke is the deck's paved layers (carriageway and shoulders) and edge_band_m, not the carriageway alone: so it covers the whole paved end.
    2. Flat ends. The stroke ends flat on the end line of the ground way that ends there (within 45° of the deck's chord), stored per end in the ceiling line's words 4 and 5 (pack2x16snorm). With the deck's own chord, 441 of 1,478 end-line points of the fixture lay past the stroke.
    3. Bank. The ceiling is the deck's banked surface across its stroke, so it meets the paved end's bank.
    4. HeldOverCap. A Held step over the paved side's cap gets its own label, which check::held_over_cap lists; EDGE_MAX does not count it. On the fixture 74 are listed, all at classes without a wall; 22 of them are small steps where the steep face would fit, so invariant 4 as written does not hold for them. Gated at 74.
    5. 2b-rail. The rail's fill less its cap joins the budget by a hard maximum (a rail bed under its cap changes nothing), and the line whose cut it lifts stands on its own wall (Cause::wall), so the rail bed's own step keeps its label.
    6. Invariant 2. 22 of 1,478 fixture end-line points are short of the drop by over 0.1 m (most 0.35 m: junction blend faces at the paved end); 197 have another paved face beyond, 46 lie under no deck. Gated.
    7. Invariant 5. 51 steep ground triangles at the corners of paved ends (the rock-face fall of an edge's cone past its end normal, which the round slope end bares). Gated at 51.
    8. EndFace is the label on the deck end line of any line present (ramps that end side by side, W1).
    9. Pictures. The 22 deck places and the five other named places; B01/p08 keep their band (the stamp's deck terms do not bind under high decks), S5 shows no opening (the heights give 4.7 m), R11 and B09 get worse. Details and sheets: the PR.

    Generated by Claude Code

  7. timohueser commented on Nov 13, 2026

    @timohueser
    OwnerAuthor

    Done: merged into ground-plan as #3396 (cda4c63). The owner saw the shots and approved: a minor improvement. Left for later steps, noted on #3353 and #3345: the opening below S5, the ground beside the deck at the Hergiswil tunnel, the hard crease at a low deck over water, the deck's flat end from the way graph.


    Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions