12. Sub-kerf degenerate-speck drop

kerfmaster pin list · UI / workflow decisions

intentmixedWhat 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 2026-08-02 — computed from the item itself, not typed.

Contract — Intake stage

This contract describes a stage we intend to build — it is NOT a description of the code.

takes raw-geometry

makes raw-geometry

fails if A feature above kerf width dropped as a speck; a sub-kerf speck left in to distort the fidelity score.

Contract last changed 2026-08-06 — computed, not typed. Dated separately from the text above, so neither date can speak for the other.

1 open question blocks this pin

Q22 When the fidelity search finds no counterpart for a point, how should a deliberately dropped speck report differently from a genuinely lost feature?
Aristide — Dropping sub-thou dirt is correct behaviour and currently reports as a fidelity failure.

Customer files can carry degenerate stray polylines — ex11 poly 2 = 3 near-coincident points spanning 0.36 thou total (20× under the 7-thou kerf → physically uncuttable). The healer rightly drops them as noise, but then the source→healed fidelity check can’t match the dropped speck and pins the score to its sentinel (ex11 read FAIL / 1e21 thou while the real geometry was a clean 0.57 thou); QC trips too.

Proposed fix: drop any polyline whose whole extent is below the kerf at parse time and log ‘N specks removed’ (never silent). Open: exact threshold (kerf / tol / own constant), drop vs. flag-and-keep, and whether a removed-speck count shows on the board.

Vector side — settled and shipping. The same defect arrives from .ai/PDF: EAGLES.ai carries four stray paths of 0.29–1.07 pt. vec_pipeline.read() both drops and reports them — any path shorter than SPECK_FRAC = 0.005 of the artwork's diagonal is excluded from the G-code and listed in meta["specks"], then ringed on the board (vec_pipeline.py:144-155; the default is keep_specks=False, and passing keep_specks=True keeps them). check_vec.py:58 applies the same gate but only reports, because it is a measuring tool rather than the pipeline. Note what the gate is not: it is a fraction of the artwork, not a function of kerf, so this item's title "sub-kerf" no longer describes it.

The DXF side is still open — and the sentinel is real

There is no parse-time speck drop on the DXF path. The healer drops degenerate segments under EPS_DROP = 5.0e-4″ (heal.py:807), silently. The 1e21 reading has a mechanism: heal.max_deviation() seeds its nearest-neighbour search at 1e18 and, if no neighbour is found within 64 rings, returns that seed without taking its square root (heal.py:911-922) — 1e18 inches, which the board prints as 1e21 thou. A dropped speck with nothing near it is exactly that case. Grepping the code for "1e21" finds nothing: the literal is 1e18, and the thou conversion makes the rest.

Pinned by Aristide 2026-07-24, found on ex11.