Skip to content

TrailView parts P6: the pair of tubes, the opening in rock and the free tube end - #3780

Draft
timohueser wants to merge 20 commits into
developfrom
claude/trailview-3683-pair-rock-free
Draft

timohueser wants to merge 20 commits into
developfrom
claude/trailview-3683-pair-rock-free

Conversation

@timohueser

@timohueser timohueser commented Nov 21, 2026 •

Copy link
Copy Markdown
Owner

Summary

P6 of the TrailView tunnel parts (#3651, #3683): with ground=plan, the viewer draws the pair of tubes, the opening
in the rock and the free tube end. Pair, Rock and Free join BUILT of obc-trailview-parts, so every chain
whose two ends have a part draws as parts. Without ground=plan, nothing changes.

The design answers of #3683 are in the issue (issuecomment-7067727023): a free end is cut square with a rim; a pair
has no middle wall where its shells overlap (the owner's option 1, made general); the ragged edge at a free shell is
the ground's.

  • The pair (tube.rs, arch.rs): a pair has an end at its mate's opening too (Net::new). Each opening draws
    the head wall of both, with its own hole, up to the plane in the middle between the walls of the two tubes. Where
    both holes reach that plane, they are one opening whose outline follows both arches; the reveal closes the step
    where the two holes differ there.
  • Tubes that meet (tube.rs, mesh.rs): where the shells of two built tubes overlap, the inside and the shell of
    each end at the middle between their walls, and faces close the steps where their floors, rooms or shells differ.
    One test, Net::meets, finds the other tube: by cells of the tile, where its chord crosses the station's line
    between the planes of its rings (at an opening its frame line, at a bend its mitre), at the level of the room. The
    middle is measured square to each tube's own chord, so both tubes find one plane at any angle. A clip holds only
    between two stations that both meet the other tube; where the other starts or ends inside a piece, halvings find the
    place, the clip ends there (a bound plane) and a face closes the shell's section. A clip takes off the side toward
    the other's axis, also where a narrow tube lies in a wide one's room. Each tube runs 5 cm past the middle, as its
    vertices are rounded to world units. The tube labels come from one union-find walk, made only for the meshes.
  • A head at another tube (arch.rs): the head of a built end ends at the middle where another built tube crosses
    its frame line, or the approach of an opening behind it does (within pair_max_m), and that tube's room overlaps the
    head from its foot to its top. A face closes the head's section there: from the foot to the top, but the hole from
    the front to the frame line and the room from there to the back.
  • The rim (arch.rs): rock and free have no head wall and no wings. A rim, the frame grown by the shell
    (0.95 m round the room), hugs the opening from the foot up, with the reveal and the frame of the arch. A free end's
    shell runs its collar from it.
  • The closed end (arch.rs, boxes.rs, tube.rs): a free end covered at its opening (collar 0, meets no way)
    has no opening: a dark face closes its tube on the frame line; it has no box in front, no ring and no daylight.
  • The ground plan (openings.rs): the crate's mouths gives each opening of a built end, two of a pair, so the
    approach of a pair's mate widens to its jamb lines too.
  • The bake (ladder.rs, mod.rs, parts.ts): a pair needs a part at both ends of both its chains (12 pairs of
    revision 35 had a chain with an end of no part, which never draws as parts); a pair's record takes the heading of the
    opening farther in, square to its frame line, along which its mate is measured. Each tube of a pair has its own
    collar (mate_collar), so its ring, shell, boxes and ground stamps end where its own hill starts. The face of a pair
    holds its band over the higher crown of the two tubes and its cap over the lower; its foot reads both collars and
    both floors.
  • Features revision 36 (obc-data-steps), the spec (rung 5, heading, collar, face, foot and
    mate_collar), the recorded sizes of the parts and plan bridges (wasm_size_guard.py).
  • Tests: the tubes of a pair, or two or three arches side by side, are one room (no face in it, every way across
    it meets the front of a face), with one opening where their holes overlap and a head wall between them where not;
    with floors 0.8 m apart; the boxes hold the room. A rim hugs its opening; rock, free and a closed end are closed
    but at their foot; a closed end has no box in front and no daylight. Two tubes meet only where both are: staggered
    openings, a tube that turns away inside a chord, a narrow tube in a wide one's room, a pair 25 degrees apart; a cut
    head is closed (a taller wall, frame lines 3 m apart). The bake: a pair whose second chain forks in the hill is no
    pair; each tube of a pair has its own collar.

Deviations from #3683

  1. The issue's size says "no new contract". The spec changes in a few lines (rung 5: a pair needs a part at both
    ends of both chains; heading of a pair; each tube of a pair has its own collar, with mate_collar; the face
    and foot of a pair read both tubes), and the bake changes (revision 36).
  2. The cut of two tubes is general, not only for a pair: any two built tubes whose shells overlap meet at the middle,
    and a head ends at another tube or at the approach of an opening behind it. This also changes develop's arches and
    underpasses where two of them stand side by side.
  3. The slanted cut of the free end is square (90°): see answer 1 of the issue.

Pictures

On the owner's PC: /media/timo/OSM1/tmp/ground-plan/parts/p6/pics/ (INDEX.md with a caption per place and
camera; sheets/ with today's app | new roads (develop) | new roads with the parts; boxb/ full size; fix/ with
the parts before | after the fix round of the review; slider.html). Box B, 13 places of the three kinds, 4 places of
develop's parts, and 9 places of the fix round (staggered and angled neighbours, former pairs), from the road 25 m and
4 m out, from above and from inside. Develop is 1f7b5f5 with features r35; the parts column is 156e769 with
features r36b (the bake of dd67223; the last commit changes only the recorded bridge size).

  • Works: one head wall for the two tubes of a pair (the prototype's faults at 717 and 642); one opening and one
    room where the tubes overlap, with no fin and no wall of one tube in the other; a cut head is closed; a tube keeps
    its wall where the other is not yet there (stagger-rail-inside: a pier between two staggered openings); the rim
    of the openings in rock and of the free ends; a closed free end is dark from inside (today's app and develop show the
    hill through it: BROKEN there); the arch and the underpass of P5 do not change. At pair-mway the trees behind the
    left part of the head wall come back.
  • Faults: at free-open a thin post stands in the one opening of a free end and an arch: the free end lies 1.4 m
    in front and their rooms overlap 3.7 m, and the free tube keeps its wall until the arch's tube starts (the review's
    rule for staggered openings). At pair-step a 2.5 m face closes the step between two track beds (data). At
    pair-mway a smaller pale shape stays behind the head wall (not the collar any more; source not found). At
    stagger-path a path tube lies in the road tube's room (data), so only part of its rim stands. At stagger-box
    specks of sky at the foot of a cut head. At was-pair-1 a grey-blue area beside the arch's opening, the same before
    the fix round. At near-box a slot of sky between two underpasses' head walls, as on develop (their shells do not
    meet). At pair-thin a dark wedge of ground in front of the pier. The dark blocks at free-cliff are a gallery's.
  • BROKEN: none that the parts make. Before the fix round, stagger-path had no road tube (sky through it).

Measurements

Develop 1f7b5f5 against this branch (bd57096 for the frame time, the first render, the cold starts, the bake and
the bridges; the pictures are of 6276fec, whose faces are the same), timosPC, GTX 1660 SUPER, beside the terrain
run and other agents (load 11 to 14): relative numbers.

Cost develop this branch
Frame, fixture walk, ground=plan (GPU ms per frame, 3 alternated rounds) 8.445 (8.410, 8.445, 8.484) 8.359 (8.321, 8.359, 8.421)
Frame, fixture valley-drive, ground=plan 8.291 (8.279, 8.292, 8.291) 8.244 (8.244, 8.205, 8.267)
Frame, fixture walk / valley-drive, today's app (1 round) 7.373 / 6.984 7.395 / 7.049
First proper render, fixture walk, ground=plan (median of 12 loads, 4 alternated rounds) 7,051 ms (first frame 1,138 ms) 7,356 ms (first frame 1,280 ms)
First proper render, fixture drive, ground=plan 7,870 ms (first frame 1,180 ms) 8,478 ms (first frame 1,347 ms)
Cold start on box B, parts ok (14 paired loads: 6 P6 places and the walk, 2 rounds, the #3773 method) 5.09 s 5.07 s (paired −0.01 s; −0.53 to +0.49 s)
Cold start: first frame / eye ready / near ready / idle 1.25 / 4.48 / 6.19 / 8.66 s 1.30 / 4.36 / 6.29 / 8.70 s (paired −0.02 / −0.01 / −0.01 / +0.17 s)
Cold start: stamp worker / parts bridge, per load 5.32 / 0.21 s 5.42 / 0.30 s (the bridge later in 14 of 14)
Box B features bake (4 threads, alternated base, head, head, base; fix round: one run of dd67223) 322 s, 277 s; peak 2.04, 1.98 GB 338 s, 402 s; peak 1.97, 2.03 GB; fix round 227 s, peak 1.97 GB
Box B feature archive 11,331,691 B 11,332,398 B
Parts bridge, CI (gzipped wasm + glue / raw wasm) 125,922 B / 293,466 B 147,785 B / 335,603 B
The net of a ground plan cell (mouths(), develop built_ends()), one tube of 100 / 300 / 900 chords, native release, median of 41, 3 alternated rounds 15–20 / 45–62 / 155–157 µs 14 / 41–42 / 141–145 µs
meshes() of two tubes of 300 chords, 7 m apart (they meet) / 30 m apart 12.9–13.1 / 12.8–13.4 ms 19.8–21.1 / 12.7–13.1 ms
Plan bridge, CI 828,795 B / 2,129,601 B (record) 831,313 B / 2,133,837 B
  • Frame time: no change beyond the noise; the same passes (13 with the plan, 11 without), nodes (130 walk, 157 drive)
    and structure pool (24 MB). The fixture holds one underpass and no end of the new kinds.
  • First render: in rounds 1 and 2 the branch was slower, in rounds 3 and 4 faster (walk 6.5 s to 7.3 s against
    6.6 s to 8.5 s); the noise of the PC is larger than the change.
  • The parts bridge does more work: the new kinds draw now, and each station looks for a tube that meets it. Per
    place, the bridge's time per load (bd57096, before the fix round): pair-mway 536 → 743 ms, pair-track 1,018 →
    1,532 ms, free-cliff 551 → 753 ms, walk 56 → 91 ms. It runs in the stamp worker; the time until the parts are right
    does not change. The fix round makes the net of a cell linear again (it was 2.3 ms at 900 chords) and keeps the cut
    at about +55 % where two tubes meet; I did not measure the cold starts again.
  • Bridge size: +21.9 kB gzipped against develop (+17 %; the fix round adds 7.0 kB); a clamp with variable bounds
    brings the formatting of floats into the WASM (+24 kB), so the parts use max and min. The ground plan bridge grows by about 2 kB, as its openings now go through the parts' net.
  • The bake of dd67223 gives the archive of the pictures (sha 2e777e0d); develop's gives revision 35 (sha 75058dcc).
    The measure reads 233 openings again (55 on develop), as all built stretches open. Ends: 130 pair openings (154 on
    develop: 12 pairs became arches; box B's tiles hold 9 of them), 327 arch (303), 191 free, 49 rock, 504 box. The fix
    round changes only pair records: all 55 gain mate_collar, 15 change collar, 21 face, 10 foot.

Checks

Run:

  • obc test -p obc-trailview-parts (12), -p obc-trailview-features (133), -p obc-ground-plan (165, before the last
    change to Net::meets, which the ground plan does not call), -p obc-data-steps (80, first round)
  • cargo clippy --all-targets -D warnings for obc-trailview-parts, its bridge, obc-trailview-features and
    obc-ground-plan; wasm32 for obc-trailview-parts-bridge and obc-ground-plan-bridge
  • cargo fmt --all --check
  • wasm_size_guard.py --module parts and --module ground-plan; the records are the CI sizes of 3ac162b
  • builder/ground-plan-wasm/check-bytes.sh: 256 of 256 cells give the native digest
  • the fixture bake with this branch: the same archive as revision 35 (sha 64b8b99b); the six fixture tiles are the same
  • the box B bake of dd67223 gives the archive of the pictures (sha 2e777e0d)
  • a ray scan of box B's 824 built openings with the crate (a way up from the room at 1, 3, 6 and 12 m in meets a vault,
    a way into the opening meets no face): against 2908211, the fix round fixes one place (stagger-path) and changes
    no other
  • vitest structures/parts (the change of parts.ts)
  • npm run check, obc prose --check, obc suites check

Left out: npm run check:wgsl and the phone run (no shader or viewer change), the browser suites and the rest of the
web vitest (CI), obc test affected (CI), the cold starts and the frame time again (no change of the viewer).

Public docs: no change.

Requirements: none

🤖 Generated with Claude Code

timohueser and others added 10 commits November 21, 2026 19:13
… its record the heading of its frame line

A pair is one part of two chains, so the bake pairs two openings only where each chain has a part at both of its
ends. The record of a pair takes the heading of the opening farther in, square to the frame line that holds both
openings, as its mate's metres are.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… free tube end, and a closed free end

- A pair has an end at its mate's opening too. Each opening draws the head wall of both with its own hole, on its
  side of the middle between the two tubes' walls; along the chains, the inside and the shell of each tube end at
  that middle where the other tube is near, and faces close the steps where the floors, rooms or shells differ.
- Rock and free ends: a rim hugs the opening (no head wall, no wings), in the end's look; a free end's shell runs its
  collar from it.
- A free end covered at its opening has no opening: its tube ends in a dark face, with no box in front and no daylight.
- All five kinds are built. The ground plan opens and widens the approach of a pair's mate too.
- The mesh keeps only the vertices that a face uses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…dle between their walls, with no gap

A tube finds the other by its cells, not by the record of its pair, so a tube in a tile with no record of its pair
ends at the middle too. A station a metre past a chord still meets it, as the two openings of a pair lie on their
frame line within the rounding of the record. Where one ring of a chord meets the other tube, the plane runs along
the chord. Each tube runs 5 cm past the middle, as its vertices are rounded to world units.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r opening that leads the same way

Not only the head of a pair: two built ends that are no pair, whose axes pass within reach of a head, and that lead
the same way within pair_max_m and pair_max_deg at one level, end their heads at the middle between their walls. So
two underpasses, two arches or a free end beside an arch side by side do not stand in each other's openings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e on each side, not only the nearest

A tube between two others ends at the middle on each side, and so does a head with a near opening on each side: the
mate of a pair on its side, else the nearest one that leads the same way. The mesh builder clips by several planes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ading is square to its frame line

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tep takes max and min, not clamp

`f64::clamp` checks its bounds and so brings the formatting of floats into the WASM (about 24 kB raw). A clip's map of
cut vertices and the tubes' cells take key and value types that the crate's maps already have.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d the cut of tubes

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

coderabbitai Bot commented Nov 21, 2026

Copy link
Copy Markdown
Contributor

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

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

Share

…anch

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

@timohueser timohueser left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Review of 2908211

Result: 2 blockers, 4 should-fix, 3 nits. The pair of two parallel tubes on one frame line is correct. The general
rules "two tubes meet at the middle" and "a head ends at a near opening" fail where the two parts do not start on one
line. I proved the faults with throwaway tests on the crate (rays from the room). The full review with numbers is on
the owner's PC: parts/p6/REVIEW-3780.md.

Blockers

  1. A cut runs on over the whole chord (tube.rs Net::clips, 382-408). Where one station of a piece meets the
    other tube and the next station does not, the clip plane goes parallel to the own chord over the full piece. The
    tube then has no wall on that side, and a slot in its vault, where the other tube is not there.
    • Staggered openings: A from 0 m, B 7 m to the side from 3 m. From A at 0.5, 1.5 and 2.5 m in, a ray to the side
      meets no face within 30 m. A rider sees out of the tube, and the sky shows through the roof. This is very likely
      the sky slot at near-box.
    • A tube that ends or turns away inside a chord: A has no wall on that side for the rest of the chord.
    • Fix: clip only between two stations that both meet the other tube. Split the piece where the other tube starts
      or ends.
  2. A head cut leaves the wall open (arch.rs head_cuts and the clip, 121-262). Only the reveal step gets a
    closing face, and only on one frame line. The wall's own section at the cut stays open where the two heads differ.
    • Two arches on one line, one wall 2 m taller: a ray inside the taller wall goes 11 m to a back face.
    • Two arches with frame lines 3 m apart: both walls are open at the middle.
    • You can see this in free-open-eye (the arch wall over the rim).
    • Fix: close the section in the cut plane, as tube::shell does with cut_face and steps. Cut a head only where
      the other part crosses its frame line, not for each opening within 40 m.

Should-fix

  1. The mate end takes the pair's collar (the larger of the two tubes). The ring of the covered tube then takes
    10 m of hill: the ground look changes and the trees go. This is the pale patch in pair-mway-top (record: collar
    10). Fix: write the collar of each tube in the record.
  2. Tubes and heads meet by two different tests. The tube cut has no heading limit, and at an angle its middle is
    off by more than the 5 cm overlap (0.23 m at 25°). Both tests ignore chains that are not built. The 12 former pairs
    are now an arch beside a tube of the old code, with no cut between them. Please show 2 or 3 of them.
  3. Cost. Net::new finds the tube of each chord with a loop that repeats until nothing changes. For a chain of n
    chords this is quadratic: mouths() takes 66 µs, 302 µs and 2.32 ms for 100, 300 and 900 chords (develop
    built_ends(): 16, 47, 161 µs). It runs per node, per tile and per ground-plan cell. Only meshes needs it. Fix:
    one walk (BFS or union-find), only in meshes. meshes() of two 300-chord tubes that meet: 19.3 ms against
    13.9 ms on develop.
  4. The cap and the foot of a pair come from the record's tube only. Over the lower tube the wall can pass the cap
    by the floor difference (pair-step: 2.5 m). The foot does not read the mate's collar ground.

Nits

  1. mesh.rs 132: "one of at most four" clips. There are at most two.
  2. ladder.rs whole searches all openings per tube. Make the (tube, end) map one time and use it for whole and
    spans.
  3. A head cut reaches 40 m along. That can be past the 40 m tile buffer, so two tiles can decide differently.

The three points in the pictures

  • pair-rail-eye. The top is flat at face 10.62 m: the ground on the frame line is that high. The slanted sides
    are the wings: they fall 2:1 to the coping in 5 m (the P4 rule for a short row). The rule is correct and the wall is
    under the cap. The wall is heavier than P5 (2:1 wings, 3.6 m over the shell), but this is the known P5 case "a wall
    on a sloped line stands high", not a P6 fault. The owner must decide on it.
  • pair-mway-top. This is the ground ring of the covered tube (item 3). It is not a shell and not a fade tail.
    Fault of this PR.
  • free-open-top. This is the shell. Its width and foot are correct. The ragged line is where the ground crosses
    the shell, so it is the ground's. The only fault of the PR there is item 2.

Not checked

I did no render, no bake and no browser suite. I tested only obc-trailview-parts (16 pass, my probes included).

🤖 Generated with Claude Code

timohueser and others added 8 commits November 21, 2026 21:40
…ube crosses the frame line

A clip holds between two stations that both meet the other tube; where the
other starts or ends inside a piece, the clip ends there and a face closes the
shell. One test (`Net::meets`) serves tubes and heads, with the middle measured
square to each tube's chord, so both find one plane at any angle. A cut head
closes its section. The tube labels come from one union-find walk, made only
for the meshes. Each tube of a pair takes its own collar (`mate_collar`).

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

The record of a pair writes `mate_collar`; the device stops the ground plan's
stamps of each way at its own tube's collar. The face of a pair holds its band
over the higher crown and its cap over the lower; the foot reads both collars
and both floors. One map of the opening at each tube end serves `whole` and
`spans`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… pair read both tubes

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

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d min, not clamp

`clamp` with variable bounds brings the formatting of floats into the WASM.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Where a narrow tube lies in a wide one's room, the middle between their
walls lies past the narrow one's axis: the clip took off the wide tube's own
room. It now takes off the side toward the other tube's axis.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d ends at an approach and over the heights it spans

`Net::meets` reads where the other tube runs from the planes of its rings
(at an opening its frame line, at a bend its mitre), not from its chord's
vertices: the two openings of a pair lie on one frame line, and their chord
vertices up to half a metre off it. A head also ends where the approach of an
opening behind its frame line crosses it (within `pair_max_m`), so it keeps
out of the way to that opening, and it meets a tube whose room overlaps its
wall from its foot to its top, so a pair whose floors lie far apart still
ends at the middle. `Meet` holds how a part looks for other tubes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…and a head's closed cut

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

Copy link
Copy Markdown
Owner Author

Fix round for the review of 2908211

Head: 3ac162b. Box B is baked again (features r36, the bake of dd67223, sha 2e777e0d). The pictures are of
156e769 (the last commit changes only the recorded bridge size): pics/fix/ on the owner's PC holds the parts
before | after this round for every shot, pics/sheets/ the three columns.

B1. A cut holds only where both tubes are (Net::clips, mesh.rs). A clip holds between two stations that both
meet the other tube. Where only one station meets it, 14 halvings find where the other tube starts or ends on the
piece; the clip holds to there and no farther (a bound plane in the mesh builder), and a face closes the shell's
section past the middle there (end_section, drawn unclipped, as it lies on the bound). Tests (rays from the room):
staggered openings, B 3 m after A: north from A at 0.5, 1.5, 2.5 and 2.95 m in meets A's wall; at 5 and 15 m B's far
wall; west from B's room into A's shell meets the closing face. A tube beside A for 20 m that turns away: at 30, 60
and 90 m a way north meets A's wall.

  • near-box: B1 is not the cause of its slot. The two underpasses there lie 4.8 m apart at 45 degrees; their shells do
    not meet, so no cut runs there. Its pictures are the same before and after this round.
  • One more fault of the cut, found in the new pictures (stagger-path, BROKEN before this round): where a narrow
    tube's axis lies in a wide tube's room, the middle lies past the narrow one's axis, and the clip took off the wide
    tube's own room (no inside, sky through it). A clip now takes off the side toward the other's axis. Test: a path
    tube 3 m from a road tube's axis.
  • Cost of the rule at one place: free-open (a free end 1.4 m in front of an arch, rooms overlapping 3.7 m) was one
    clean opening; now the free tube keeps its wall until the arch's tube starts, and that wall stands as a thin post in
    the joined opening (pics/fix/free-open-eye.png). The alternative, a clip that also holds along the approach of
    the other's opening, gives back the one opening there, but also the open side of your staggered probe. I left the
    rule as you asked; this is a call for the owner.

B2. A head ends only where another tube crosses its frame line, and its cut is closed (arch.rs). The head cut
is Net::meets at the frame station, with the reach of the head's wall. A face closes the head's section at the cut:
from the foot to the top, but the hole from the front to the frame line and the room from there to the back. Tests:
two arches on one line, the first's wall 2 m taller: over the lower wall a way to the taller one meets the closing
face at the middle (7.6 and 8.6 m up). Frame lines 3 m apart: both heads end at the middle, closed (ways at 6.2 m up
meet them 3.5 m out), and the way into the second opening meets no face. Three things the box B pictures and a ray
scan showed, each with one rule in meets:

  • A tube is met between the planes of its rings (at an opening its frame line, at a bend its mitre), not between its
    chord's vertices: the mate's chord vertex of a pair lies up to half a metre off the frame line, and with 10 cm the
    head of pair-over stood over its mate's hole.
  • A head meets a tube whose room overlaps the head from its foot to its top, not within half the two heights: three
    pairs on box B have floors 3 to 5 m apart, and their heads were not cut.
  • A head also ends where the approach of an opening behind its frame line crosses it (within pair_max_m, the reach
    of a pair). Without it, the front head's wall stands across the way to the opening behind: at one pair of
    underpasses 0.25 m apart on box B (46.86576, 8.63483) it covered the whole opening. This goes past your wording
    ("only where the other part crosses its frame line"); the reach across is the head's own width, so N3's 40 m across
    is gone, and the reach along is that of a pair.

S1. Each tube of a pair has its own collar (ladder.rs, mod.rs, parts.ts, spec). The record of a pair writes
mate_collar (one row of the tag table; the collar line says each tube has its own). The mate end takes it, so the
ring, the shell and the boxes of each tube end where its own hill starts, and the viewer stops the ground plan's
stamps of each way at its own collar. Bake test: under a ridge that rises half as fast over the second tube, the
fork test's pair has collars 4 and 9 (west) and 7 and 12 (east). On box B all 55 pair records gain mate_collar; 15
change collar. At pair-mway the left tube's collar is 4 m (was 10): the trees behind the left part of the head
wall come back. A smaller pale shape stays; the collar is no longer its cause, and I did not find its source.

S2. One test for tubes and heads (Net::meets). The middle is the point on the station's line where the metres
to the two walls, each measured square to its own tube's chord, are equal: both tubes find one plane at any angle, so
the 0.23 m offset at 25 degrees is gone; I added no heading limit. Test: a pair whose tubes run 25 degrees apart from
one frame line: ways north from the first meet the second's far wall, and ways up anywhere between the two axes meet
a vault. Limit: two openings at an angle that are no pair, each with its own frame line, whose rooms overlap: a
corner of sky can open between the one's cut and the other's frame line. Box B has no such place (2 angled
neighbours, 7 and 12 m apart, no overlap).
Former pairs: box B's tiles hold 9 of the 12 (64 → 55 pair records). 4 have a built arch beside a tube of the old
code (near-arch, was-pair-1 to was-pair-3), 5 have both tubes in the old code (was-pair-4). Fine: no fin and
no gap from the meeting of the two. At was-pair-1 a grey-blue area shows beside the arch's opening; it is the same
before this round (the old code's tube lies in the arch's room), not investigated.

S3. The tube labels: one walk (Net::meeting). One union-find over the vertex map; the labels and the cells are
made only in meeting(), which only meshes calls. Native release, median of 41 runs, 3 alternated rounds:

chords mouths() this branch built_ends() develop
100 14 µs 15–20 µs
300 41–42 µs 45–62 µs
900 141–145 µs 155–157 µs

meshes() of two tubes of 300 chords: 19.8–21.1 ms where they meet (7 m apart) and 12.7–13.1 ms where they do not
(30 m), against 12.9–13.1 and 12.8–13.4 ms on develop. The clip costs about +55 % where tubes meet, nothing
elsewhere. I did not measure the cold starts again.

S4. The cap and the foot of a pair read both tubes (ladder.rs, spec). The wall holds its band over the higher
crown of the two and its cap over the lower; the foot reads both collars and both floors. On box B the face changes
at 21 pair records (for example 11.5 → 6.1 m at two rail pairs whose floors lie more than 5 m apart, 7.3 → 9.1 m
where the mate's crown is 3 m higher), the foot at 10. The rail pair's head wall (pair-rail) keeps its height.

N1. The comment says one clip per side and four planes per key; a debug assertion holds it. N2. One map of the
opening at each tube end serves whole and spans. N3. Settled by B2: the reach across is the head's width.

Checks. obc test -p obc-trailview-parts (12), -p obc-trailview-features (133), -p obc-ground-plan (165, before the last change of meets, which the ground plan does not call);
clippy for these and both bridges (native and wasm32); fmt; both size guards; check-bytes.sh (256 of 256); the
fixture bake (same archive, same six tiles); vitest structures/parts; npm run check; obc prose --check; obc suites check; a ray scan of box B's 824 built openings (a way up from the room meets a vault, a way into the opening
meets no face): against 2908211 one place fixed (stagger-path), none worse. Left out: browser suites, cold
starts, frame time, the phone.

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

@timohueser timohueser left a comment

Copy link
Copy Markdown
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Delta review: 2908211..e9e4f30

The full review is in DELTA-3780.md next to REVIEW-3780.md on the owner's PC. I proved each finding with throwaway ray probes in obc-trailview-parts (not committed). obc test -p obc-trailview-parts passes at head (12).

Verdict: 1 blocker, 2 should-fix, 5 nits. I recommend one rule in the bake instead of patches on the device (see the end).

Blocker

  • D1. A head cut where its own tube is not cut leaves the shell open at the front (arch.rs:121-131, tube.rs:441). Rule (c) cuts the head where the approach of an opening behind it crosses the frame line. The own tube is not cut there, so its wall and vault past the middle stand on behind the head, and no face closes their section at the frame line. Probe, B 3 m behind A: a ray west from inside A's wall meets nothing. With B 20 m behind, a ray from in front into A's wall meets nothing within 30 m. This matches the specks of sky at stagger-box-near, and very likely the front of the post at free-open and of the pier at stagger-rail.

Should-fix

  • D2. B1 still misses overlaps that start and end between two stations (tube.rs:485-525). A 6 m tube inside one 10 m chord of A: A's shell stands 2 m into its room. Two tubes in a row beside A with a 3 m gap: in the gap, A has no wall and no vault (rays north and up meet nothing).
  • D3. The approach is the first chord, straight for 40 m, at any heading. A tube 22 m away at 45 degrees cuts A's wing, where its real way may never pass.

Nits

  • D4. end_section lies on the bound only if the station is square to the chord. T-junctions are possible.
  • D5. meets/clips give no cuts without meeting(). A debug assertion would catch that.
  • D6. The spec's pair face line does not say which bound wins (the code keeps the band). The line is not wrapped.
  • D7. The head's level test reaches down to the foot, so a tube under a portal at an angle can cut the head.

Sound

mate_collar agrees in the spec, the bake, the device and parts.ts. The cap and foot over both tubes are right. The union-find is right, and only meshes calls meeting(). The closing faces of two heads on one line face the right way. The clip side for a narrow tube in a wide room is right. Rule (a) is right for a pair.

free-open, and one rule for all three owner notes

There is no rule on the device alone. A post (or pier) and an open side are the two answers to the same shape: one room beside the other's open approach. The rule belongs in the bake: where the rooms of two tubes overlap, they are one room.

  1. Absorb: a tube whose axis lies in a wider tube's room for more than half its length has no part. The wider tube widens, and the way runs in it (stagger-path). This replaces today's same-tube test at one opening.
  2. Join: if the rooms overlap at an opening, the two openings are a pair, for any kinds, on the frame line of the opening farther in (free-open, stagger-rail, stagger-box: one head, no post, no pier).
  3. Meet and part in the hill: the bake puts a vertex where an overlap begins and ends. The device cuts only pieces whose two rings meet.

This fixes D1 to D3. The halving, the bound plane and the approach leave the device: about 75 lines fewer on the device and about 80 more in the bake. Pairs that work now do not change. I did not implement or bake it.

Public docs: not reviewed for change; the spec rows match the code.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant