The build spec, glossary & backlog — one source of truth · updated 2026-09-21 · click any item for full detail
By subjectBy pipeline stageBy evidenceOpen questions
python3 pinlist/record_debt.py for the list.Single source of truth: cadmaster/pinlist/pins.json, edited live via the editor (✎ on any item). This page and PINLIST.md are generated from it — version-controlled and pushed to GitHub on every save.
Vocabulary — Heal stage
Fidelity = how closely our healed output matches the customer’s TRUE original shape. The single number that answers ‘did we change the part?’
So 0.57 thou means the healed curve never strays more than ~half a thousandth of an inch from where the real edge should be. The laser kerf is ~7 thou, so 0.57 is well inside the beam — invisible on the cut part.
Why it’s the priority. Everything downstream (tabs, leads, nest, G-code) is only as good as the geometry going in. If healing silently reshaped the part, we’d cut the wrong thing perfectly. Fidelity is the proof that input = output before we touch anything else. The PASS/FAIL banner on every comparison board is exactly this: best-fit fidelity in thou vs true source, plus the jog count.
Explained for Aristide 2026-07-24.
Vocabulary — Cross-cutting stage
Thou = one thousandth of an inch (0.001"). Our working unit for tolerances and fidelity. Reference points: healing gate 10 thou for these parts (down to 1 thou for real work), best-fit target ~0.5 thou, laser kerf ~7 thou. So a fidelity of ‘0.57 thou’ is about 1/12 of the beam width — invisible on the part.
Vocabulary — Cross-cutting stage
Kerf = the width of material the laser beam removes as it cuts (~7 thou here). It sets the floor for real detail — anything narrower than the kerf can’t be cut. Two consequences: (1) the post applies cutter comp (G41/G42) with a D-offset (kerf .007 → D7) so the beam centre is offset half the kerf and the part comes out on-size; (2) geometry smaller than the kerf (a 0.36-thou stray speck) is physically uncuttable noise — see item 12.
Vocabulary — Heal stage
Jog (a.k.a. kink) = a non-tangent break between two segments — a visible corner where the true shape is smooth. A fit-induced jog (one our healer introduced) is an automatic FAIL even if it’s within tolerance: it’s a defect we created, not something in the customer’s file. Only true source corners stay sharp. The verifier checks both within-tolerance AND zero fit-induced jogs; the comparison board draws a red circle on any it finds. Contrast spike vs sustained.
⚠ Known defects
This item's subject has a known defect in the code. Recorded and deliberately NOT fixed — recording a defect is not a licence to fix it. Source: cadmaster/defects.json → cadmaster/KNOWN_DEFECTS.md, not this page — the text below is generated from the register so the two can never disagree.
Vocabulary — Heal stage
How the healer decides corner vs curve at a vertex:
The robust test is neighbour consistency, not raw turn size: repeated ~equal turns at ~equal spacing = arc to smooth; an isolated big turn between straight runs = corner to keep. Genuinely borderline chains get flagged for a human, not silently guessed.
Where it lives in the code (2026-08-03): heal.py — CORNER_DEG 35.0 (the turn that may be a corner), ALWAYS_DEG 55.0 (always a corner), CONSISTENCY 2.5 (the spike multiplier — how much bigger than its neighbours a turn must be), CORNER_MIN_FACET 0.002, JOG_DEG 1.0. heal() returns corners, borderline and corner_pts in its meta, and that is exactly what "flagged for a human" means today: a count and a coordinate list, not a prompt.
Vocabulary — Heal stage
Bulge = the DXF encoding of a native arc inside a polyline: bulge = tan(¼ × included angle). Sign gives direction (+ CCW, − CW); the centre sits on the minor-arc side of the chord. These are the customer’s real, intended arcs — we preserve them exactly, never re-fit them from sampled points. (Re-fitting is only for curves the customer exploded into straight facets — see polylining.)
Vocabulary — Heal stage
Polylining = exploding a real curve into hundreds of tiny straight line segments. It’s what dirty CAD (and generic PDF/AI converters) do to circles and arcs — a circle becomes 100 little chords. It’s the exact damage cadmaster exists to undo: we re-fit those facets back into true G2/G3 arcs. We must never produce polylined output ourselves — a circle stays a circle.
Vocabulary — Heal stage
Healing = turning a dirty faceted file into clean geometry — re-fitting straight-line facets into true arcs, keeping real corners sharp, dropping degenerate micro-segments — all within tolerance and with zero fit-induced jogs. The core of what cadmaster does. Measured by fidelity; the corner/curve call is driven by spike vs sustained.
Healed geometry is the output of that process: the clean set of lines and true arcs that replaces the customer's faceted input. Everything downstream works from the healed geometry, never from the raw file — QC, islands, bridges, leads, nesting and the post all read it.
On the comparison board the word is literal: <label>_healed.dxf is the cut geometry alone, and <label>_healed.nc is that same geometry as motion (inspection only — see item 4a). The customer's original is never overwritten; it stays as the gray base layer underneath.
A file is only called healed once it has passed both gates: within tolerance against the TRUE source, and zero fit-induced jogs.
A distinction CLAUDE.md makes non-negotiable, and this item never made. A DXF is healed: facets are re-fitted into arcs, and the measurement is deviation from the true source geometry with bulges expanded. An .ai/PDF is fitted: true cubic Béziers are converted into arcs, and the measurement is a different one — reported three ways as the raw deviation, the format's own inherent error at that radius (a cubic Bézier cannot be a circle; Illustrator's is out by 0.02725% of radius), and what our fit adds on top. Only the third number is ours. The two boards therefore carry different pass/fail measurements on purpose (items 2c, 25).
Vocabulary — Heal stage
The tolerance argument is an accept/reject GATE, not a fit target. We always best-fit as tightly as the geometry allows (target ~0.5 thou, FIT_TARGET=5e-4); the tol passed to heal() (e.g. 0.010") is only the line above which a result FAILs. So a 10-thou gate does not mean we fit to 10 thou — we fit far tighter and use the gate to reject anything that couldn’t get under it.
Vocabulary — Intake stage
Layer = a named grouping that every entity in a DXF belongs to. CAD packages use layers to switch groups of geometry on and off, colour them, and decide what is real geometry versus notes. We use them in three distinct places, and they should not be confused with each other.
A customer's layers may carry meaning we need (cut vs engrave vs construction lines vs dimensions) or may be meaningless. We never assume. Related: item 8 (cut vs engrave) and item 10 (export layer table).
Cut geometry and annotation are kept strictly apart, so a file can go to the machine without anything to explain away:
HEALED — the healed cut geometryORIGINAL — the customer's geometry, for referenceCM_BRIDGE_OK — bridges measured at or above the table minimumCM_BRIDGE_ADJUST — undersized bridges, with the proposal drawnCM_SPACING_FLAG — narrow webs that are spacing, not bridgesCM_NOTES — text calloutsThe shop's DXF standard is R2000-2002 (AC1015), so that is what we write. We also ship a mark-up-only DXF containing the annotation layers and no geometry, for opening alongside the cut file.
The cut-path review DXF is a second table, and it is not this one. cutpath_dxf.LAYERS writes CUT, LEAD, PIERCE, SEQ, DIR, RAPID, RAPID_HAZARD, HELD, LEGEND, TAB, TIP_ARM, TIP_CENTRE. Both tables exist in code today; item 10 lists neither in full.
House style: the original DXF is always the base layer, drawn thick and gray, with healed and generated work overlaid on top and each overlay toggleable. That is what makes a discrepancy visible instead of arguable. Half met, 2026-08-03: the vector board honours it (srclayer, fitlayer, hazlayer groups, each with a checkbox — build_vec_board.py:215-217, 364-372). The DXF board does not: the source paths are written straight into the SVG with no group id (build_board.py:195), and the only checkbox on that page toggles the bridge / spacing mark-up (brlayer). See item 6.
⚠ Known defects
This item's subject has a known defect in the code. Source: cadmaster/defects.json → cadmaster/KNOWN_DEFECTS.md, not this page — the text below is generated from the register so the two can never disagree.
Vocabulary — QC stage
Island = a piece of kept metal fully surrounded by cut-away area (the MVT play triangle inside its ring; the centre of an ‘O’ or ‘A’; any stencil counter). Cutting the surrounding hole frees it and it drops out — so an island needs bridges to stay in place — permanent, by design (item 2b). A tab is the sacrificial hold, at a different stage (items 3a, 3b). Detected geometrically in bridges.islands(): the material is rastered even-odd at ~20 thou/px, every candidate neck is cut in the raster, and connected components are labelled once — a region that separates only because of those cuts was being held by them, so it is a real island and those necks are its bridges. Full spec: item 2b. Assembly caveat (2026-08-06): connectivity is a property of the assembly, not of the artwork — a shape can look joined in the composition and be a free island the moment the plates are split, so the island test is re-run per plate after every split. See item 2d.
Vocabulary — Nest stage
Tab = a connection that is meant to be broken. Its entire purpose is that the material comes apart there easily, later in the manufacturing process. A tab is a production feature, not a design feature — it holds cut parts to the surrounding sheet (the skeleton) so they don’t drop or tip into the machine during the cut, and it gets broken out afterwards. Nothing about a tab is intended to survive into the finished part.
Tab width/thickness values are set at nest time and vary by material and thickness — see item 3a.
⚑ There are three holds, not two. A bridge is permanent and stays in the finished design (item 2b). A tab — this item — is sacrificial and holds the part to the surrounding sheet at nest time (item 3a). A slug-retention tab is sacrificial too, but sits on the waste side, holding an interior drop-out so it cannot fall through the bed (item 3b). Our spec originally used "tab" and "bridge" as synonyms — that was wrong.
Which of the three the code emits today: only the slug-retention tab. emit._slug_tab_width() sizes it and emit._leave_uncut() leaves the gap, on interior boundaries, at post time. The nest-time tab described in this item needs a position on a sheet; there is no nester, so nothing places one.
MetaCam’s word for this kind of sacrificial joint is wirejoint (Jordan, 2026-08-28).
src — Jordan, 2026-07-28 (terminology correction: tab = breaks off, bridge = stays).
Vocabulary — Bridges stage
Bridge = a connection that stays. It is part of the overall design: bridges are what hold islands in — the centre of an ‘O’, the counters in text, floating elements of art — so those pieces remain attached in the finished part. A bridge is intended to be there, and intended to be seen.
The exact opposite of a tab, which exists precisely so it can be broken off. Bridges are permanent; tabs are sacrificial.
Full spec and pipeline placement: item 2b.
src — Jordan, 2026-07-28 (terminology correction: tab = breaks off, bridge = stays).
Vocabulary — Post stage
A slug-retention tab is a short length of an interior boundary deliberately left uncut, so the waste piece (the "slug") stays put instead of dropping through the bed or tipping up into the head.
Confirmed by Jordan, 2026-07-29, after it was measured out of the shop's own production NC for the novi sign: every one of the 90 interior boundaries stops a constant 0.0181" short of its own start point, and the outer profile closes exactly. The customer DXF contours are all closed, so this is added by the post — it is not inherited from the file.
All three post the same way — a gap left in the cut path — and all three are sized for different goals. A lead or pierce must never land inside any of them.
Answered by Jordan, 2026-07-29: it scales with material thickness, and the values will be defined in the tables. So this is the third width table alongside bridges (item 2b) and tabs (item 3a), with the same treatment — ships as defaults, lives in operator Preferences, editable per laser and per location.
Superseded by a real test cut, 2026-07-30 (Aristide) — see item 3b for the numbers. The headline: the driver is slug size and leverage, not thickness alone. The "0.30 × thickness" ratio is dead: 0.018" held on 0.060 material and held again on 0.125 material for small slugs, where a ratio would have demanded 0.037". A big slug on the same 0.125 sheet needed 0.033", and a slug under 3/4" needed no tab at all. So the table is keyed on material × thickness and slug size.
Failure is not cosmetic: an undersized tab lets the slug drop early and the loose piece can hit the laser head. When a slug falls between two bands, round up.
Vocabulary — Nest stage
Wirejoint = MetaCam's term for the sacrificial micro-joint — a short span left uncut so material stays joined until it is deliberately broken. In our vocabulary that is the tab (part held into the sheet, item 3a) and the slug-retention tab (waste side, item 3b). It is never a bridge, which is permanent.
src — Jordan, 2026-08-28 ('1. Yes.'), confirming the term from the Add Toolpath walkthrough step 1: MetaCam's Auto-tool settings dialog carries a 'Suppress wirejoints' checkbox, CHECKED in the shop's standard run — MetaCam's automatic joints are off in their flow. How tabs then get placed (manually per feature?) is expected later in the walkthrough — not yet asked.
Vocabulary — QC stage
Undersized = a bridge, tab, slug-retention tab or any other narrow web that measures narrower than the table minimum for that material and thickness. It is a measurement against our table, not a judgement about the customer's file.
CM_BRIDGE_ADJUST layer — the wider web and a tick at each wall's new position — and a human decides.91 contours, 92 narrow webs: 73 real bridges holding 46 islands, 0 unbridged. Of those, 10 measured under table — 59/63/64/73/74/75/75 thou in text against an 80 thou minimum, and 113/115/121 thou on the big art island against 125. A further 9 narrow webs under 80 thou were spacing, not bridges, including the 23.3 thou U–T gap, which is the narrowest of the nine (23.3 / 34.3 / 52.8 / 58.1 / 65.9 / 67.6 / 68.0 / 72.2 / 79.0). Re-measured 2026-08-02 at HEAD 0ded91a; the 10 under-table widths are 59.0 / 63.3 / 64.4 / 73.0 / 74.2 / 75.0 / 75.5 in text and 113.5 / 115.1 / 120.6 on the art island, all at the board’s 0.060 thickness setting.
Undersized is not the same as defective. It means our table wants more material there than the file has, on this material and thickness. Change the material and the same geometry may be fine.
⚠ Known defects
This item's subject has a known defect in the code. Source: cadmaster/defects.json → cadmaster/KNOWN_DEFECTS.md, not this page — the text below is generated from the register so the two can never disagree.
Vocabulary — Starts & leads stage
Lead-in / lead-out = the short ramp on and off the cut path. The pierce (where the beam plunges through the sheet) sits on the lead, off the part edge, so the pierce scar isn’t on the finished contour. A lead / pierce must never land inside a tab span. Choosing lead position, direction, and kerf side (G41 vs G42) per contour is item 4.
Vocabulary — Nest stage
Nesting = packing multiple copies of a part onto a stock sheet for best yield (least waste). Base stock 4×8; materials copper / aluminum / stainless / mild steel in various thicknesses. Must honour the bed envelope. Item 3.
Two kinds (Jordan, 2026-08-27): true nesting (same part duplicated — toolpath first, then nest) and part-by-part nesting (several unique shapes, Torongo files — nest first, then toolpath).
Vocabulary — Nest stage
True nesting = the same part duplicated over and over on the sheet (Jordan, 2026-08-27: 'When the same part is duplicated over and over (we call this true nesting)'). The LS1156258 384-piece sheet and the 3A76 lattice sheets are true nests.
Pipeline order: toolpath FIRST, then nest. The single part is tooled completely — leads, cut order, slug tabs — and the nest duplicates that finished toolpath: 'that way we use one short piece of GCode duplicated over and over.' This is why MetaCam's array spacing measures the toolpath bounding box and why the flip-every-other-column trick aligns pierce points (confirmed by Jordan same day) — what gets arrayed is a toolpath, not bare geometry.
Contrast: part-by-part nesting. Full ruling and consequences: true vs part-by-part nesting.
Vocabulary — Nest stage
Part-by-part nesting = several unique shapes nested together on one sheet (Jordan, 2026-08-27: 'We also have nesting that applies to several unique shapes (Torongo files) where nesting would happen before the toolpath is established. We call this part-by-part nesting'). Torongo files is the shop's name for these multi-part jobs.
Pipeline order: nest FIRST, then toolpath. The geometry is laid out on the sheet, and leads, cut order and tabs are established afterwards across the assembled sheet — the reverse of true nesting. The split_parts multi-part layouts are this kind.
Full ruling and consequences: true vs part-by-part nesting.
Vocabulary — Nest stage
Bounding box = the smallest upright (axis-aligned) rectangle that fully encloses a part — its width × height on the sheet. Cheap to compute, so it is a useful first-pass estimate for sizes, envelope checks and rough layout.
Ruling (Jordan, 2026-08-26): “For nesting, gaps will be between actual part geometry, not the bounding box.” The ruled part-to-part gap (the Rev2 min-gap column) is the closest approach of the parts’ true outlines. A part that doesn’t fill its rectangle — like the BNM78651 V-strip — leaves huge dead space if boxes are spaced instead of outlines.
Consequences: every nester slides parts to true-outline clearance, and every nest audit measures the min gap on densely sampled real geometry — a bounding-box spacing check can only ever be a fast pre-filter, never the gate.
⚠ Known defects
This item's subject has known defects in the code. Recorded and deliberately NOT fixed — recording a defect is not a licence to fix it. Source: cadmaster/defects.json → cadmaster/KNOWN_DEFECTS.md, not this page — the text below is generated from the register so the two can never disagree.
Vocabulary — Cross-cutting stage
Bed envelope = the machine's cuttable area: 80" × 160". Anything larger physically cannot be cut on this laser.
⚑ Not checked everywhere, 2026-08-03. This item said we check every file, single part and full nest. What runs today: the vector path checks it (vec_pipeline.envelope_check, called from build_vec_board.py). The DXF board never checks it at all — build_board.py prints the extents and never compares them to the bed; 80×160 appears in that file only inside the unit-detect heuristic. There is no nester, so no nest is checked. And the check that does run silently ignores the machine profile: envelope_check reads machine.get("bed_x") / ("bed_y") while emit.MACHINE_DEFAULT stores "bed": (80.0, 160.0), so the hardcoded 80×160 always wins and a different machine would be ignored without a word. Recorded as a known code defect; deliberately not fixed here.
Vocabulary — Post stage
Fanuc macro-B = our production G-code dialect. O-number program, condition variables (#102–#111 plus #121/#122 — counted in the master NC; #112–#120 do not exist there, and the pinned "#102–#122" implied twenty variables where there are thirteen), G41/G42 cutter comp with D-offsets, M98 subprogram pierce calls, G52 work shift, true G2/G3 arcs with I/J. The 0.2 margin is applied at post via G52 X.2 Y.2 — never baked into DXF coords. (This item read Y.95 until 2026-08-02; the master NC is and always was G52X.2Y.2, re-verified from the file itself.) The comparison board hands back both a geometry .nc and this production post.
⚠ Known defects
This item's subject has a known defect in the code. Source: cadmaster/defects.json → cadmaster/KNOWN_DEFECTS.md, not this page — the text below is generated from the register so the two can never disagree.
Vocabulary — Post stage
Jordan, 2026-08-04, asked directly because the word was blocking a fix: "A subprogram will be made from a dxf file, using the same gcode multiple times repeated throughout a nest. (A single gcode can contain the same string multiple times. Example, if we want to draw a circle 100 times on the same sheet, we'd have 1 main program calling out 1 subprogram 100 times.)"
So a subprogram is one part — the whole of a part's cut, all its features, authored once from one DXF. The main program is the sheet: it calls that subprogram once per instance in the nest. This is not the same thing as the pierce subprograms (O9011 and the rest of the 9-series), which are Mitsubishi firmware we call but never author.
It resolves what fires M121. Jordan's rule is "M121 at the end of every subprogram", which now reads: at the end of every part, not at the end of every feature and not once per sheet. Against the master NC that is consistent — one part, 91 boundaries, the last of them ending M121 where the other 90 end M87. A 100-up nest of that part would carry 100 M121s, one per call. See defect D13.
cadmaster has no part-level subprogram. Every boundary is written inline into one main program; the only M98 we emit is the pierce call. That is correct for a single part and it is what the master does, but it is not the structure a nest needs. When nesting is built (item 3), the post has to learn to emit a part once and call it N times — and the natural mechanism for placing each call is the work shift, G52, which is the same command we currently use for the 0.2 margin. Not built, not designed, recorded here so the shape is known before anyone starts.
Both structural questions were put to him: should a single part be emitted as a real subprogram even when it is called once, and is G52 the mechanism that positions each call on the sheet. The same answer to both: "Keep it this way for now, and we'll learn more about subprograms and nesting later."
So the inline structure stands, and this is a deliberate hold rather than an oversight. The consequence for D13 is that its fix is the small one: with everything inline, the subprogram is the program, so M121 goes on the last boundary and M87 on every other. No restructuring rides along with the M-code fix.
⚠ Known defects
This item's subject has a known defect in the code. Source: cadmaster/defects.json → cadmaster/KNOWN_DEFECTS.md, not this page — the text below is generated from the register so the two can never disagree.
Vocabulary — Report stage
Comparison board = how every result is delivered — one hosted page at cadmaster.aristide.mycloudally.com with the original DXF as a thick-gray base layer, healed lines/arcs overlaid, red circles on any jog, download buttons (healed DXF, geometry G-code, Fanuc post), and a PASS/FAIL banner with fidelity + jog count. Built with build_board.py. The board is the artifact; the Telegram message is just the pointer.
Where it lives, since 2026-08-02: the site root is now the numbered build history (item 5). Every build keeps its own page at builds/NNN-<label>/ and nothing is overwritten. A vector board (item 2c) is a different page reporting different measurements, and takes a number in the same sequence.
Shop DXF — the standard shop-bound export (D29, closed 2026-08-07): every build also emits <label>_shop.dxf, the same healed geometry re-spelled as R12 bulge polylines + true circles (dxf_compat.py default), and it is the primary DXF button on the board. MetaCam mis-draws standalone ARC entities as full circles but reads the bulge spelling perfectly — it is the customer corpus’ own diet. A bulge is an exact arc: G-code posted from either spelling through the same emit.write_gcode() agrees to 0.04 thou, under half of one count of the 4-decimal NC format (nc_from_dxf.py is the proof harness). The R2000 _healed.dxf remains for CAD that reads it (SolidWorks-verified).
Vocabulary — Cross-cutting stage
Workbench = the docked ground state of the interface picked 2026-08-19 (A+C, item 7b): canvas owns the screen, the eight workflow tabs across the top wear status colours, the left toolbar follows the active tab, panels sit in fixed dock slots. Every station boots into the same Workbench, so any operator can sit at any seat. It is the part that ships FIRST, and it ships looking designed — the Graft machinery underneath is a fast-follow on a day-one architecture.
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — Cross-cutting stage
Graft = the tear-off machinery layered on the Workbench: any docked panel can be torn off to float, dragged to a second monitor, and snapped back to its dock slot in one click. The Graft is what makes per-user workspaces possible. It is architected in from day one (retrofitting a dock-only UI is a rewrite) but switched on as the fast follow after the docked Workbench ships.
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — Cross-cutting stage
Workspace = a named, per-user saved arrangement of panels, Adobe-style: which panels are open, docked or torn off, on which monitor. Switching workspaces is one click; snap-back to the default is one click; the default workspace is the Workbench itself. Workspaces are remembered per user, so a shared station follows whoever logs in.
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — deliver stage
Floor packet = the paper (or PDF) sheet that travels to the machine. Floor operators generally do not use cadmaster; they run the program from this packet — nest preview, material, program number, material usage, and a time estimate when one exists — and make machine-side changes on the control itself. It is a Deliver-stage pipeline output, one packet per program. Spec: item 7c.
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — Intake stage
Quote mode = the stripped-down sales surface: upload a customer file, assign material, auto-nest, get material-use and time estimates, price with the per-shop minimum-fee table. A saved quote becomes the starting point the operator programs from — "it moves the start, it doesn’t move the program" — and customer approval is an explicit gate before production finishing. Spec: item 7d (risky-flagged; nothing built).
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — Cross-cutting stage
Interrupt-and-return = the first-class ability to pull a job back. A job that has already been programmed or delivered can be brought back in, changed, reposted, and the operator resumes where they left off — losslessly, with nothing re-entered. Ruled first-class at the sitting (not an edge case): shop reality is that jobs get interrupted, and "our software wraps itself around their workflow".
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — Cross-cutting stage
Gate vocabulary = the designed visual language in which cadmaster says ready / attention / blocked — the status colours on the workflow tabs, dots, glows and flags on geometry. It is a vocabulary we OWN and teach once, in the first-run tour (item 7f), not something diluted to resemble every other tool; the bar is iPhone-grade — a visual learner reads it in minutes. The mechanical form of a hard block (modal vs flag) is still open and follows the vocabulary.
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — Cross-cutting stage
Units are a skin, never a second source of truth. Display units are a per-user preference (Canada wants metric), and every numeric field accepts a unit suffix, Illustrator-style: type 3mm at an inch station and it converts on entry. Internal math stays inches; the 4-decimal emit floor (4d-4) is unchanged. Spec: item 7e.
src — design sitting, Aristide + Jordan, 2026-08-19; full record at /design/answers/.
Vocabulary — preflight stage
Ruling (Aristide, 2026-08-10): “we need to make a distinction between logo’s / artwork and regular files.” The MVT logo work is parked on his word; Jordan is feeding regular (non-logo) files that need cleaning — these define the regular class by example (derive-don’t-ask).
What the classes are FOR: which rules bind. Artwork is judged by the eye — tolerance follows what a viewer can see, smoothing within the gate allowed, connectivity/bridges matter. Regular files are dimensional — holes take fasteners, edges mate, sizes bind. One PART can carry both classes as features (the Brazen tread: functional profile + mounting holes + lettering), so the classification is ultimately per-feature-function (Jordan 2026-08-06), with the file class setting the default.
First regular-class file (record): DWS-LH-2.5-LOGO-TREAD-BRAZEN-PS.dxf, 2026-08-10 — 17.375×3.25 in tread, $INSUNITS=1, 43 LINE + 9 ARC + 6 CIRCLE + 45 cubic SPLINE on named layers (Outer_Loop / Interior_Loops). CAD-native export, not polylined art: true splines, declared units, semantic layers. Healed PASS at 1.60 thou / 0 jogs (build 012).
Second regular-class file (record): the DS twin, 2026-08-10 — same tread, different lettering. Caught a real defect: a source ARC meets a 0.023 in whisker spline at an 8.35° junction; the fit added 1.79° (10.14° emitted) and correctly FAILed. Two causes fixed: eval_spline's 64-sample floor turned the whisker into sub-mil micro-arc noise, and the Q27 smoothing pass ran before chain_contours, so an entity-to-spline junction was never an adjacent pair and could not be blended. Fix: smoothing now runs after chaining (heal.py); DS heals PASS 1.53 thou / 0 jogs / 14 junctions blended, worst blend 1.4 mil vs the 5 mil budget. PS output byte-identical before/after. Corpus check unchanged by the fix — but it exposed pre-existing drift at HEAD, logged as D30.
Ruling (Jordan, 2026-08-27), verbatim: 'When the same part is duplicated over and over (we call this true nesting). For true nesting, the nesting process should come after the toolpath is established, that way we use one short piece of GCode duplicated over and over. We also have nesting that applies to several unique shapes (Torongo files) where nesting would happen before the toolpath is established. We call this part-by-part nesting.'
True nesting is why MetaCam's Layout Array spacing measures the TOOLPATH bounding box (leads included) and why the flip-every-other-column trick aligns pierce points — the thing being arrayed is a toolpath, not bare geometry. Raised as my flagged inference from the two 2026-08-27 screenshots; Jordan ratified it same day, so it now reads as agreed fact.