Aixle

Appearance

Everything is done: only dogfood/ and out/ were touched, nothing is committed, and the final artefacts are in out/sailboat/ (model.glb 9.6 MB, model.stl 15.3 MB, beauty.png 1024 px, viewer.html). Report follows.

Dogfood report: sailboat ​

1. The model ​

A hard-chine sailing dinghy, 5.4 units long over the rudder (hull 4.9), 6.9 tall to the masthead, on a wooden two-frame cradle. Bow is +z. Program: dogfood/sailboat.aix (240 lines, 54 named steps, all in the output).

How it is built:

  • Hull: three station tables (keel, chine, sheer as z, half-beam pairs) are turned into symmetric polygons by loops, then two stacked lofts (keel→chine 0.3 tall, chine→sheer 0.8 tall) make a V-bottom with a hard chine and a raked stem and transom. The curved sheer is a big cylinder (r 11) lying along x subtracted from the top; rocker is an intersection with another (r 12) resting on the keel. def sheer(z) is the deck height used by every fitting.
  • Cockpit: a rounded trapezoid polygon extruded and subtracted; a thwart across it.
  • Paint: hull painted gloss navy, then four decals: deck (the sheer cylinder grown by 0.03), antifouling (everything under the waterline), white boot-top (a 0.1 slab), and the well last.
  • Gunwale: a tube through 24 deck-edge corner points, each found with surface() from a guess just outside and above the corner.
  • Rig: mast, masthead sphere, a masthead crane with anchored port/starboard tangs, gooseneck collar; boom is a joint at the gooseneck carrying the boom, a boom-end fitting (anchor end) and the mainsail. Sails are extruded triangles (0.12 thick; the main with a 4-point roached leech) given a belly by wrap, turned with rotate(y=-90), and moved so the luff lands on the mast axis; the jib is additionally tilted by the forestay's rake, computed with atan from the stem fitting and masthead anchors, with the clew given in world coordinates and projected along/across the luff.
  • Rudder: a joint "helm" on the stock at the transom: stock, blade, rudder head, tiller; two gudgeons on the hull.
  • Rigging: forestay (masthead→stem fitting), two shrouds (crane tangs→stern-quarter posts), a bridle between the quarter posts, the mainsheet from at(boom_rig, "end") (so it follows the pose) to the bridle's middle, and a jib sheet from the clew to a deck fairlead. Every line is a tube between anchors.
  • Cradle: two beams and two rails on y = 0, four posts leaning in from the beams to surface() points under the bottom panels, with rubber pads.
  • Poses: rest; sailing = boom 35° about y, helm −20° about y; animation tack (rest→sailing→rest, 3 s).

check at rest (grid 320, cell 0.0216, no warnings): hull 1.81 × 1.08 × 4.89, deck edge from 1.80 amidships to 2.06 at the transom and 2.09 at the stem; mainsail 0.31 × 4.13 × 2.03 (belly 0.2 to port); jib 0.24 × 4.22 × 2.8; boom 2.09 long at y 2.57; masthead at 6.82; sailboat 2.2 × 6.9 × 5.38. check --pose sailing: boom_rig posed x −1.38..0.07 (end at x −1.15, z −1.0), tiller posed x −0.46..0.05, mainsheet rebuilt 1.21 wide to follow the boom end.

2. Print readiness and re-import ​

Final out/sailboat/report.md (grid 320, cell 0.022, 305,392 triangles, volume 4.65):

RowValue
Pieces1
Watertightno: 37 edges. Mostly 4 at (−0.04, 2.599, 0.334) and 3 at (−0.041, 2.601, −1.13) in main_flat/boom: the sail's inner face leaving the boom's top (a sub-cell wedge where the 0.12 cloth, bellying away, crosses the 0.045 boom); 3 at (−0.604, 2.016, −2.241) in upper_hull/sheer_cutter: the deck cut meeting the loft's side at the stern quarter. The rest are single edges on the shroud tubes (0.07 = 3.2 cells). Down from 98 at first; the remaining ones are surface crossings, not thin parts.
Standsyes, centre of mass (−0.02, 1.98, 0.04) is 1.06 inside the 20-sided footprint
Overhangs13% (the V-bottom panels and the sails' undersides; a boat on a cradle needs support under the hull)
Cavitiesnone; rests on y = 0

Re-import (dogfood/sailboat_reimport.aix, import("../out/sailboat/model.glb", resolution=200), 73 s to sample 314,368 triangles): size 2.201 × 6.897 × 5.386 vs 2.2 × 6.9 × 5.39 — identical; Pieces 1; Stands yes; volume 4.55 vs 4.65 (−2%, the 0.034 sampling cell shaves the thin parts); watertight no (79 edges, mostly around the rudder stock at cell 0.034). Everything survived: both sails with their belly, all six rigging lines, gunwale, tiller, rudder blade, centreboard, cradle posts; the boom and rudder are at the right heights, so the GLB's boom/helm nodes (translations (0, 2.566, 0.6) and (0, 2, −2.47)) are placed correctly. Materials are not read, as documented.

3. Friction (every stop) ​

  1. loft is centred on y = 0, the docs say y = 0 to h. sailboat.aix line 64 (lower_hull = loft(keel_p, chine_p, chine_y)). check printed lower_hull 1.52 × 0.3 × 4.6 y -0.15..0.15; docs/language.md says "loft(a, b, h) blends from profile a at y = 0 to b at y = h". Expected y 0..0.3. Fixed with | move(0, up + chine_y / 2, 0). Everything downstream (sheer cut, rocker, the gunwale's surface() points all at y 0.7) was wrong until then.

  2. A white part lit face-on is invisible on the sheet — I spent ~9 renders "finding" a sail that was never missing. Lines 167 (mainsail), 190 (jib). From the first full render the RIGHT and PERSPECTIVE views showed the jib grey and, where the main should be, only background with a faint leech line; steps.png showed the sail in steps 28/29, the Y = 3.45 slice showed both arcs, --focus mainsail showed it, the OBJ/STL had 56,370 vertices in its box. Repainting it "maroon" (sailboat_probe_dark.aix) showed it in place. The ivory cloth (#f6f2e6) with its normal towards the camera renders in the sheet's background value; the tilted jib picks up shading. Nothing in the docs' "mistakes the renders catch" table covers a part that is present but reads as absent. Worked around by never trusting a white part's absence again.

  3. The quick sheet mislaid a decal, and I drew the wrong conclusion. Line 95 (hull = ... | decal(well_region, ...) last). Render 1 (--quick, cell 0.108) showed the cockpit floor red though the well decal was applied last; I "fixed" it by reordering, and the full render then showed red for a real reason. Probes (sailboat_probe_decal.aix, sailboat_probe_well.aix) established: the last decal wins, and at a 0.108 cell the floor's vertices sat below a region that ended 0.03 under the floor. The docs say nothing about decal precedence or about vertices straying out of a region on a coarse mesh. Fixed by keeping the well last and extending its region 0.1 below the floor (line 94).

  4. surface() from a far guess slid to the wrong place in a loft's field. Line ~230 (p = surface(hull_cut, x, up + 0.1, z)). With the guess at y −0.5 the point came back at [0.13, 0.04, 0.94] (near the keel) for x = 0.7; the posts converged on the keel in the FRONT view. With a guess 0.1 under the panel it returns [0.5, 1.3, 1] — exactly the chine height, 0.07 above the true panel. The reference says "exact on primitives and close on blends and warps"; a loft is neither and the result is only "close". Worked around by guessing near, and the 0.09 pads hide the error.

  5. A transform above a joint is not carried into the GLB's joint nodes. Old line sailboat = (boat + cradle) | ground(). The GLB had boom at translation (0, 1.566, 0.6) and helm at (0, 1, −2.47) while the hull mesh was grounded (+1); the OBJ had the blade at y −0.68..0.57 next to a cradle at 0..1.27. The sheet and report were right; only the exports were wrong, silently. Fixed by building with an up = 1.0 offset in every y and no ground() (line 240). Docs say "exports are at rest" but not "and ignore transforms above joints".

  6. at(shape, "centre") on a joint is the box centre, not the part. Line 214. I wrote at(boom_rig, "centre") meaning the boom end; the mainsheet ran from mid-air at y 3.6 and the watertight report blamed mainsheet. My error, but the report's location (−0.016, 2.424, −2.16) was what caught it, not the sheet. Fixed with anchor("end", ...) on boom_end (line 139).

  7. The wrapped luff does not land where the flat luff was. Lines 149–150, 166, 189. Overshooting the luff by lap moves the wrapped edge ~lap·sin(half/r) (0.017) off the axis, so a 0.03 stay in 0.1 cloth left jib_flat/forestay edges at x = 0.044. The reference for wrap gives no formula for where a point lands. Fixed by moving by the angle halfway along the overshoot and thickening the cloth to 0.12.

  8. A stay leaving a sail at a corner made a "piece". Line 180. Render 13 reported Pieces 2 (... 0 at (0.064, 5.776, 0.996) in 'jib_flat', 'forestay') — a zero-volume speck at the jib's head. Fixed by running the luff to the masthead so the exit is inside the masthead sphere. A zero-volume speck counting as a piece is noise the report should filter.

  9. Watertightness of rigging. Line 205. Tubes at 1.6 cells (r 0.022) gave 98 edges; 2.2 cells still 44 on the tubes alone; 3.2 cells clean. The docs' "1.2 cells" threshold is for being caught at all, not for being watertight; nothing says what watertight needs.

  10. Two shrouds meeting at one point leave a sub-cell wedge. Old masthead line. Report: "8 at (−0.047, 6.339, 0.297) in 'shroud_p', 'shroud_s'". Fixed with a crane and two tangs (line 131). Not documented.

  11. check's surface line missed thin plates. boat surface ... y 0..5.9 z -2.62..2.54 while the centreboard reaches −0.9 and the rudder blade −2.84; both were in the mesh. The line is described as the "true extent".

  12. The steps sheet's "in output: yes" and the report do not say whether a step made it into the mesh, only into the tree; nothing distinguishes "present but white" from "dropped".

  13. Command-line friction, not the tool's: the sandbox refused every heredoc/sed pipeline as "too complex", so probe files had to go through separate scripts.

Budget: 15 renders of the model (quick, 7 full, 2 focus, 3 grid trials, beauty candidate, pose sheet) plus 13 renders of probe programs and the re-import. Beauty at 1024 took 18 s; a full grid-320 render with exports about 60 s (hierarchy 13 s, anim 8 s).

4. What worked well ​

  • check in under two seconds, printing numbers as well as shapes: the whole jib geometry (rake, cu, cv, the clew position) was verified before any picture.
  • Anchors and at() for the rigging: six lines fitting-to-fitting with no arithmetic, and at(boom_rig, "end") following the pose so the mainsheet stays on the boom end in sailing.
  • The watertight report naming the two steps and the point: every remaining edge was diagnosable from that line alone.
  • --focus with its own cell, and the slices drawn from the field, which are what finally proved the "missing" sail was there.
  • wrap for sail bellies, surface() for the gunwale corners (24 points, all right), loft for a hull that reads as a hull.
  • The re-import loop closing on the GLB, which is how the joint-placement bug and its fix were both measured.

5. Three changes that would have helped most ​

  1. Make a present-but-invisible part impossible on the sheet: draw silhouette edges darker and/or shade by normal so a white face-on surface differs from the background; and let steps.png or the report say per step whether it produced triangles in the final mesh (a step "in output" that yielded no faces is exactly what I feared). This alone would have saved ten renders.
  2. Fix loft to match its doc (y = 0 to h) or fix the doc, and add to docs/language.md: the last decal wins; a transform above a joint must not be used (or make the exporter apply it); at(x, "centre") is the box centre; surface() is only "close" on a loft.
  3. Print-readiness guidance in cells: state that a tube needs about 3 cells across to be watertight and that two surfaces crossing at under ~45° or a stay leaving cloth at a corner will leave edges; and drop zero-volume specks from the Pieces count while still listing them.