kerfmaster pin list · UI / workflow decisions
recordmixedreads as factStates what the code does today. Auditable against the code — this is the only type that can go stale.
Content last changed 2026-11-19 — computed from the item itself, not typed.
Jordan's ruling 2026-11-04: fix the healer for ANY upload, never per file; anything under the heal tolerance is irrelevant to the toolpath; views stay in Source mode. heal.drop_sub_tol_pieces runs inside heal() after every source piece is fitted and BEFORE close_micro_gaps / chain_contours. A whole source piece (stitched polyline, LINE, ARC, CIRCLE, SPLINE, ELLIPSE) is dropped when (a) it is shorter than the heal tolerance and is a dot (ends within CHAIN_WELD) or not a connector (both ends must join other pieces' endpoints to be kept), or (b) its WHOLE path lies within gate-minus-smoothing-budget (0.005 at default) of other kept ink (a copy), but only if dropping it leaves no other piece with a loose end (guard: at each end of the copy, count other kept piece-ends; exactly 1 = the copy is the only joint, so it stays). Every drop is listed in meta['subtol_dropped'] (pt, len, why=dot/scrap/on-ink); not yet shown on screen. ROOT on 480125: the source has two one-segment polylines 6e-9 in long; heal._dedup keeps a run's last point by design, so a 2-point zero-length run survived as a LINE; one stitched into the corner (branch flag), one stayed alone and became a 1-segment closed 'profile' with its own lead-in (review contour 7). With the dots gone, the corner spline drawn twice (once each way) showed as dangling + branch; rule (b) drops the copy. Results 2026-11-04: all 5 rear-radar uploads (1303-1 x3, RH x2) fresh-upload 1 flag -> 0 flags; 1303-1-1793804560 review profiles 2 -> 1. First version of (b) dropped 7 short grill corner pieces (0.012-0.016 in) on 78598 and opened 7 outlines (identity gate caught it: 2 -> 8 warnings); the loose-end guard fixed it, gate green. Rule test test_subtol_pieces.py 8/8. test_healedit.py fails the RJB-DS check on BOTH old and new code (stale 273-prim expectation; heal gives 173 on both; not O89). RAMIFICATIONS: every upload door that heals (fresh uploads and re-heals; stored jobs re-heal on next heal). Could break: a real connector or thin feature under 0.010 in (connector rule + guard protect joints); two parts sharing one drawn edge (Jordan's point 2, his answer cut off -- whole-path rule means a closed part outline is never dropped, only a separate piece lying entirely on other ink); auto Drop duplicates finds fewer rows because the healer already dropped copies (those drops are not operator rows -- they are below tolerance by ruling). STILL TO DO before dev testing: finish all-parts sweep (115 dev parts, old vs new) and look at every changed part; F1 bench before/after; 44558b-7 time/memory; restart (inflight 0 both sides + ask). Supersedes the patch-style O87 heal/QC fix (tag wip/o87). || 2026-11-04 16:04 UTC: first all-parts sweep slice showed 15 of the 23004 drawing-sheet parts going heal PASS -> FAIL (fidelity 0.0046 -> 0.0152 on 23004-344): a copy judged within the full 0.010 gate, plus up to 0.005 of later smoothing on the line it was kept against, read 0.015 against the source. Fix: copies are judged within gate minus the smoothing budget (0.005 at the default gate; Q27 one end-to-end budget). 23004-344 back to PASS (fidelity 0.0051). Identity gate green; test 8/8; 480125 still 0 flags, 1 profile, PASS. || 2026-11-04 16:09 UTC sweep slice after the band fix (20 parts compared, all 23004 drawing sheets): all 19 that passed still pass, flags down on every one (2099 -> 1961 total). OPEN: 23004-798 (already FAIL on old code): inside a ~0.01 in cluster of tiny drawing-sheet marks, dropping two 0.0058 in scraps turned one 101 deg jog into two 33 deg jogs -- a scrap whose far end sits near but not ON other ink acts as a connector; not yet decided. Remaining: rest of the 115-part sweep, F1 bench (needs a quiet box), 44558b-7 time/memory, then restart for Jordan to test. || 2026-11-04 16:35 UTC FULL SWEEP (115 dev parts, fresh-upload old vs new, no errors either side): 72 parts changed; no part went PASS->FAIL; one part went FAIL->PASS (cp-6005-03-1791891707); total open-end flags 7267 -> 6843; no part gained flags. Only regression: 23004-798 (already FAIL) jogs 1 -> 2 (the open item above). Clear wins: all 480125 rear radar 1 -> 0, the 448/perf family 1 -> 0 (about 30 parts), b7787 13 -> 0, 44558b-7 8 -> 4, qclink 2 -> 0, 78598 grill sheets fewer flags and jogs 5 -> 3. 44558b-7 alone on a quiet box: 67.4s -> 67.5s, peak 4.25 GB -> 4.24 GB (no cost). SPEED: first version cost +0.75s on the F1 click; the drop step's lookup grid was rebuilt (cell size follows the part, long pieces screened by box, setup only when needed); results unchanged on an 11-part spot check (identical geometry). F1 first click 8.56/8.59/8.56/8.55s -> 8.81/8.83s, repeat 8.28/8.30/8.32/8.31s -> 8.53/8.53s (+0.25s / +0.22s, under the 1s / 0.3s stop lines). Identity gate green; test_subtol_pieces 8/8. NEXT: restart dev engine (inflight 0 both sides + ask) for Jordan to test. || 2026-11-04 16:36 UTC dev engine restarted on Jordan's 'restart' (inflight 0 on :8474 and :8475); engine active, app answers 200. In Jordan's testing. -- 2026-11-05 00:15 UTC FOUND: test_upload_cancel goes 15/15 -> 11/15 with O89's heal.py (dev tree; main heal 15/15; isolated in a temp copy). The 'heavy' upload is no longer in flight when the test cancels it; cause not traced (timing vs behaviour unknown). Must be understood before O89 ships.