2b. Bridges for floating islands

kerfmaster pin list · Big capabilities

intentmixedd4What we want to build. Never claimed the present, so it cannot be stale — and must never be read as a description of the code.

Content last changed 2026-08-09 — computed from the item itself, not typed.

⚑ Terminology fixed 2026-07-28 (Jordan). This stage is about bridges — connections that stay as part of the design and hold islands (text, art) in. It is NOT about tabs. A tab is the opposite: a connection made to be broken off easily later in manufacturing, applied at nest time to hold parts in the sheet — that is now item 3a. The spec previously used the two words as synonyms, which was wrong.

The problem. When a cut has an interior island — a piece of kept metal fully surrounded by cut-away area (the MVT play triangle inside its ring; the center of an ‘O’ or ‘A’; any stencil counter) — cutting the surrounding hole frees the island and it drops out. Bridges are the connections that hold the island to the surrounding metal so it stays in place — permanently, as part of the finished design. This is a real pipeline stage — placement below.

⚑ Needs Jordan’s shop knowledge to turn into real numbers — flagged below. (Aristide 2026-07-24: fold the stage into the pipeline now; the shop numbers are just the table values that fill the stage, they don’t block wiring it in.)

⚑ Build state, 2026-08-03 — this item had no status line at all, unlike 2a / 2c / 3b, so it read as an entirely unbuilt spec. About half of it ships. Built: the neck finder (bridges.find_necks), the island / bridge test (bridges.islands — raster the material, cut every candidate neck, label components once), width bands (bridges.suggest), the single decision path (bridges.review_marks), the CM_BRIDGE_OK / CM_BRIDGE_ADJUST / CM_SPACING_FLAG / CM_NOTES review layers, the review DXF and the mark-up-only DXF, and the board's bridge section. Not built: bridge GENERATION (sections 4 and 5 below), the angle optimiser (3a / 3b), a BRIDGES output layer, and post-time bridge gaps. bridges.py:21 says it plainly: "Nothing here edits geometry: it measures and reports." emit.py never imports bridges.

0. Where it sits in the pipeline

Full order: in → unit-detect → heal → QC topology (incl. island / connectivity detection) → BRIDGE GENERATION → lead-in/out + cut-order → nest → post.

• Runs AFTER QC — but correction, 2026-08-03: island detection is not a QC-topology output. qc.analyze returns open contours, dangling ends, branches and duplicates — there is no island map in QC. Islands are found inside bridges.islands(), which does its own rastering. The dependency is real, the ownership was wrong.
• Runs BEFORE lead-in/out + cut-order — a lead-in / pierce must never land inside a bridge span, and the sequencer has to treat a bridged contour as one that leaves gaps. So bridges are fixed first, then leads and order are laid around them.
• Runs before nesting — a bridge is a per-part internal feature, independent of where the part lands on the sheet, so parts carry their bridges into the nest already baked in.
• Inputs: healed closed cut-contours + the QC island map + job material & thickness. Output: per island, a bridge set {position along contour, width, count} on a proposed BRIDGES layer, merged into the cut path at post as omitted spans.
• Operator override stays — proposals draw on the comparison board (like QC flags), editable before post; same accept/edit flow as QC_FIX.

1. Island detection (automatic)

Built, in bridges.islands(): raster the material with even-odd parity (not winding — a different rule, and the right one here) at ~20 thou/px, cut every candidate neck in that raster, then label connected components once. Anything that separates only because of those cuts was being held by them, so it is a real island and those necks are its bridges. Anything not joined to the main body even before the cuts is an island the file does not hold — an unbridged island. A plain donut hole doesn't flag; a floating positive shape does.

2. Bridge width — base table × modifiers, calibrated by real test cuts

Do NOT fake a pure-physics number. Strength ≈ width × thickness × material-strength; the load is the island’s size and leverage.

• Base table: material (Stainless / Steel / Copper) × thickness → base bridge width. Seed from Jordan’s numbers, refine with test cuts.
• Thickness modifier: thicker → narrower bridge (strength comes from the depth); width ∝ ~1/thickness.
• Scale modifier: island mass grows with size², bridge strength only with width — so the absolute bridge grows a little on a bigger part, but as a fraction it shrinks. Aristide’s example: the logo at 6" wide on ⅛" steel needs a visibly wider bridge; scaled to 60" wide the bridge shrinks to the narrowest that still holds, for looks.
• Two clamps: a floor (smallest bridge the beam/kerf can cleanly leave and that survives handling) and a ceiling (widest before it looks clumsy).
• Count: small compact island → 1 bridge; long or heavy island → 2+ spaced bridges so it can’t sag or pivot.

3. Bridge placement — structural, but respect the silhouette

• Never on a defining corner/tip. A bridge at the triangle’s point blunts the shape and reads badly (Aristide’s explicit point).
• Prefer a long flat or gently-curved edge, near the island’s center of mass, bridging the shortest gap to solid metal (shorter = stronger and less visible).
• Balance the supports — e.g. two edges or the base of the triangle — placed symmetrically so it looks intentional; mirror on symmetric shapes.

3a. Consistent bridge ANGLE across the whole piece — Jordan's rule

Bridge placement is a document-level decision, not a per-island one. Jordan: find a line or angle from the artwork and use it for the bridges throughout the piece. Example: a text run mixing round and straight letterforms — place the bridges on the ‘Q’ island at 30°, and every other bridged letter (‘P’, ‘O’, ‘A’ …) gets its bridges at that same angle.

Why it matters: bridges are permanent and visible, so the piece must read as intentional. A set of individually-optimal bridges at scattered angles looks like damage; the same angle repeated looks like design. Consistency beats per-island optimum.

⚑ This supersedes pure per-island optimisation as the primary rule. Section 3’s structural guidance (shortest span, near centre of mass, never on a defining corner/tip) still applies — but as the tie-breaker within the chosen global angle, not as a free choice per island.

Algorithm sketch: (1) derive a dominant orientation from the artwork (text baseline, dominant straight edges) or take the operator’s angle; (2) for each island, cast the bridge along that angle and take the crossing that best satisfies the structural rules; (3) where the global angle gives a bad result (span far too long, or it lands on a defining tip), fall back by a defined rule rather than silently picking a different angle.

⚑ Answered — see 3b below for the decided angle rule.

3b. Angle rule — decided

⚑ First, a definition the rule needs. "Bridge at 30°" is ambiguous: it can mean (a) the bearing — where the bridge sits around the island, measured from the island’s centre, or (b) the crossing direction — the heading of the bridge span itself. We build (a): the angle is a bearing from the island centre, because that is what makes ‘Q’ and ‘P’ agree visually. The crossing direction is then whatever is perpendicular to the stroke at that point — always the shortest, strongest, least visible crossing.

Where the default angle comes from (Jordan: we decide per file, operator adjusts)

Auto-derive one angle per file, show it to the operator, let them override:

• Find the artwork’s dominant axis — for text this is the glyph slant (the italic stem axis); for upright text it is vertical.
• Set the bridge bearing off that axis so bridges cross strokes near-perpendicular.
• Evidence this is right: in the Novi sign the client’s own bridge/neck angles cluster tightly at 155–165° — they followed the font’s italic axis without being told to. Deriving from the artwork reproduces what a good shop already does by eye.
• 30° is not a house default (Jordan: it was only an example). There is no fixed house angle — it is per-file.

Symmetric shapes: same angle — CONFIRMED

Same bearing everywhere. No mirroring (Jordan, confirmed 2026-07-28). Parallel bridges give the piece one consistent "grain", which is what makes it read as intentional. Where an island needs two bridges, place them 180° apart on the same bearing — balanced and still parallel.

One angle per file, chosen by optimising across ALL islands — CONFIRMED

Jordan: "Per file, calculate the best angle. A human will ask for adjustments if necessary." This replaces the per-island fallback cascade previously proposed here, and it is a better formulation: rather than derive an angle from the dominant axis and then patch the islands it fails on, search the angle space and score each candidate against every island in the file at once, then take the winner.

• Why it is better: an angle that breaks one island scores badly, so the optimiser avoids it automatically. The "what if it doesn’t work on one island" problem dissolves into the objective function instead of needing escape hatches.
• No automatic per-island deviation. One angle, applied everywhere. We do not silently bend the rule for a difficult island.
• Scoring ingredients (per island, at a candidate bearing): span length at that bearing (shorter = stronger and less visible); penalty for landing on a defining corner or tip; penalty for high local curvature (prefer a flat or gently-curved edge); proximity to the island’s centre of mass.
• Protect the worst case, not just the average — an angle that is great for nine islands and terrible for the tenth is a bad angle. Weight the worst-scoring island heavily.
• Seed the search from the artwork’s dominant axis (for text, the glyph slant) rather than sweeping blind — in the Novi sign the client’s own bridges sat at 155–165°, the font’s italic slant, so that neighbourhood is where the optimum tends to live.

Operator loop: report the chosen angle and the islands that scored worst under it, so a human can see where to look and adjust. Surfacing the weak spots is not the same as silently deviating — we still cut one consistent angle unless a human says otherwise.

When we touch bridges, and when we do not (Jordan, 2026-07-28)

• Client file already has bridges → leave them alone. Their positions are their design decision. We measure them and say nothing further — unless one comes out narrower than the table minimum, and then we flag it and suggest an adjustment. We never silently re-cut a bridge that is within table.
• Client file has bare text / unbridged islands → we calculate both. Positions and widths are ours to choose, by the angle rule above and the width table below.
• The minimum applies to any narrow material web, not only to bridges. The gap between two adjacent letters is held to the same floor: on the Novi sign the web between the U and the T of “CUTTING” measures 23 thou, well under any sane minimum on .060 stainless, and that gets flagged and a fix suggested exactly like an undersized bridge. This is why the bridge-vs-spacing classification does not gate the report: everything under the floor is reported either way. It only decides what we propose — widen the bridge by moving the cut ends, or open the letter spacing.

The width table lives in Preferences (Jordan, 2026-07-28)

The numbers ship as defaults, not as constants. They belong in the operator Preferences panel, editable per laser and per location — what one shop has dialled in will not match another's machines. Same for the tab table (item 3a). Starting point offered and accepted as a starting point only, for .060 stainless: small text counters 0.080–0.100, medium islands 0.100–0.125, large islands 0.125–0.150, hard floor = material thickness.

src — Jordan 2026-07-28 (angle is per-file, we pick the default, operator adjusts; 30° was only an example). Same-angle/no-mirroring and one-optimised-angle-per-file CONFIRMED by Jordan 2026-07-28. The bearing-vs-crossing definition is Claude’s and still needs Jordan’s nod.

4. Workflow — auto-propose, human confirms

Software auto-detects islands and proposes bridges (position + width from the table), draws them on the comparison board like QC flags, and lets the operator nudge / resize / add / delete before posting. At post time a bridge is simply a gap left in the cut contour — the beam skips that span, so the metal there is never severed.

5. We GENERATE bridges — decided

Settled (Jordan 2026-07-28): the software generates the bridges; it does not merely detect missing ones. If the customer’s art has a floating island with no bridge, we create one. So bridge-generation is a real feature to build, not just a QC check.

• Delivered as its own layer — bridges go out on a dedicated layer so they are visibly separable from the customer’s original geometry. Ties into the layering scheme (item 6) and the DXF export layer table (item 10).
• The client reviews and approves before it is cut — we are modifying their artwork, so the sign-off is theirs. This is exactly an accept gate in the pipeline model (item 6a), and it is the one gate that faces outside the shop.

⚑ Open: exact layer name (BRIDGES?), and what the client actually receives for review — a marked-up DXF, a PDF/PNG proof, or a link to our comparison board?

⚑ Needed from Jordan (shop numbers)

• Base bridge widths for Stainless / Steel / Copper at common thicknesses.
• The minimum bridge the laser can cleanly leave (kerf/beam), and the maximum before it looks bad.
• Whether bridges get dressed/finished after — affects how visible we can let them be, and where we hide them.

src — raised by Aristide 2026-07-24 (MVT-ICON play-triangle island); terminology, the consistent-angle rule, the generate-and-client-reviews decision, the leave-client-bridges-alone rule and the Preferences-owned width table from Jordan, 2026-07-28.

⚑ Related, 2026-08-06: a bridge is the deliberate version of a hold. Item 2d covers the accidental version — metal held together only because the composition overlaps it, which disappears when plates are split. Both are measured the same way, by weld chord; only one was designed.

⚑ OPEN CONFLICT (recorded 2026-08-09, needs Jordan/Aristide): the written rule for a two-bridge island — 180° apart on the same bearing, never on corners — disagrees with Aristide’s own worked answer for a compact rounded island (the MVT option-A file, 2026-08-06): the triangle there is held by THREE bridges ~120° apart, each ~0.22" wide, each sitting just OFF a triangle corner — short radial spans the written rule would forbid. For compact rounded islands the rulebook and the owner’s worked answer cannot both be the expected outcome. Until ruled: offer both layouts and say which rule each follows; do not silently pick.