โ† back to pin list

480125 outer reads OPEN: a zero-length segment at a join (Jordan, 2026-11-03)

kerfmaster pin list ยท UI / workflow decisions

questionmixedUnresolved, waiting on Aristide or Jordan.

Content last changed 2026-11-04 โ€” computed from the item itself, not typed.

The outline is geometrically closed, but the healed result keeps a line about 0.000000006 in long at (317.73,-20.80), the same spot where auto Drop duplicates removed one doubled arc. QC counts 4 ends at that point ('branch: 4 ends meet'), so the loop reads open. Weld is the doubled-ink tool (needs 2+ features), so it cannot help, and the zero-length piece is too small to click. Proposed fix: the healer drops any segment shorter than 0.0001 in (below our 4-decimal resolution) and re-joins its neighbours. Heal-path change: identity gate, F1 timing, corpus sweep. Awaiting Jordan's go. UPDATE 2026-11-03 evening (Jordan: 'nothing visible ... something is broken, find it'; then 'Test it'): ROOT CAUSE on 480125 rear-radar 1303-1/RH: the source corner has a zero-length polyline AND the corner spline drawn twice (once each direction). Dev today keeps the zero-length polyline as a 6-millionths line -> QC 'branch: 4 ends meet'. With the O87 healer drop alone, the healer joins one spline copy into the neighbour arc and the other copy lies 100% on it within 0.006 but covers only 50% of it, so qc.find_duplicates (mutual coverage >=55% both ways) missed it -> dangling + branch. TESTED FIX (km-o87-wt, uncommitted, NOT on dev): qc.find_duplicates also flags a piece >=95% on another (DUP_CONTAIN 0.95; only ADDS pairs, mutual pairs unchanged), so the upload auto-drop removes it. Fresh-upload simulation: old code 1 flag (reproduces Jordan's), new code 0 flags, heal PASS 0 jogs, corner = one arc. Identity gate green. CENSUS of today's dev geometry: 8,666 open-end flags on 209 parts; 1,838 of them on 118 parts sit on invisible geometry (a piece shorter than the heal tolerance, or a piece lying on other ink); 7 huge files not checked. Still to run: fresh-upload old vs new on those 118 parts, all-parts change count, F1. Then Jordan's go before dev. || 2026-11-04 STATE (O87, tag wip/o87, not on dev): (1) heal.py drops a whole polyline that never leaves EPS_DROP of its first vertex -- the rule standalone LINEs already had. (2) qc.find_duplicates: a piece >=95% on another (DUP_CONTAIN) counts as a duplicate even when it covers little of the other -- the mutual-coverage rule missed a short piece on a long one. (3) server dead-ink re-check after a drop indexes only cells near the removed ink -- byte-identical output (44558b-7 geometry md5 equal), 77.3s -> 73.8s. Fresh-upload test, 30 parts: rear-radar x4 and 448-fork x3 1 flag -> 0, heal pass, 0 jogs; 44558b-7 8 -> 4 flags, 66s -> 74s (the drop now happens); 23004 drawing sheets x20 unchanged (their flags are real drawing-view lines); 78598 colorpick/bendline proofs 324 -> 322 -- NOT a healing fault: 7 colour drawings stacked, colour-pick question unanswered, so every colour cuts. Identity gate green. Earlier '35% slower' was measured under cProfile (inflated) -- corrected. Still to do before dev: remaining ~84 parts (grill pieces), F1 first click.