kerfmaster pin list · UI / workflow decisions
recordmixedStates what the code does today. Auditable against the code — this is the only type that can go stale.
Content last changed 2026-11-05 — computed from the item itself, not typed.
2026-11-04 19:30 UTC, list grill-with-bend-lines-original-1793818808: Jordan picked nearly the whole drawing and added it to row 53. (1) run_pladd/_pladd_raw wrote row 53's source.dxf as 61 MB, 1,079,495 vertices (contours of 32,249 vertices each), from a 516 KB original (86 ELLIPSE, 70 SPLINE among its entities; cause not traced). The re-heal of that part ran 450 s; the dev engine was restarted 19:38 on Jordan's OK to stop it. The donor rows' files had already been rewritten (moved out) at 19:30:34; the door never finished, so NO undo step was journalled -- the newest undo step is the 19:00 'Merge identical copies', so Undo would undo the wrong thing. The list is left half-done; a fresh upload is the clean way out. (2) No cancel button for a long-running door (Jordan). (3) Picking from the customer's original: works on the Part list Drawing view and on an unsplit upload's Source mode, but not in Source mode on a split part (srcRaw() is null when srcOrig() shows; the shift-band falls to the healed-geometry band). Against ruling same-tools-every-file. -- Proposed order (to Jordan 2026-11-04): 1 Add-to-row runaway (trace where the points come from: add-to-row or a shared reader; real curves; plus size guard), 2 picking on the original (now: Part list only), 3 cancel button for long doors, 4 healer double-line fix (grill-bendlines-doubled-edges), 5 part-32 heal crash (TypeError: cannot unpack non-iterable NoneType). Jordan asked which are patches: the size guard is a backstop, not a root fix; he keeps it. -- 2026-11-04 21:19 UTC trace (O91, /tmp/o91_repro.py, read-only on parent 1793818760): the 70 splines alone give 1,711,013 straight pieces (20 at 32,248 = the 32,249-vertex contours in row 53), the 86 ellipses 50,081; lines/arcs are fine (one piece each). Cause: server._parsed_to_contours samples splines/ellipses with step=0.005 in raw file units (mm here) and emits them as LINE chords; the normal upload road (heal.py) fits splines into true arcs instead. Also used by run_plnew, so step 2 would hit it too. Proposed fix: fit them with the healer's own spline/ellipse fitter in the file's units (curves become arcs), plus Jordan's size guard. -- 2026-11-04 21:33 UTC BUILD (O91, dev): server._pladd_fit_curves (heal.py's spline/ellipse road: _thin_samples/_dedup, true-curve tangents, fit_biarc_run at heal.TOL, steps converted inch->raw via the lineage affine) replaces the 0.005 raw-unit chords in _pladd_raw; _parsed_to_contours(curves=False) for that road. _pladd_size_guard(root, bytes): cap max(10x original drawable, 2 MB), RequestError in words, nothing written; run_pladd now builds everything first and writes donors + target only after the guard (donor flips/job.json deferred too); run_plnew guards before its writes. Measured on temp copies (JOBS_DIR swapped, live jobs untouched): grill curves 1,711,013 + 50,081 pieces -> 148 + 172 (226 arcs); worst deviation vs an independent dense sample of the true curves 0.00968 in (gate 0.010); small add (42 free curves + 1 moved) 1.9 s, 10 KB, intake done; guard-forced stop left target+donor files and donor lineage byte-identical; whole-grill add to row 1 24.3 s, 77 KB, 46 donors emptied (named back, same as before); whole-grill New part (copy) 40.0 s, 75 KB, intake done; overlay plot checked by eye. Tests: identity gate green; test_split_door 17/17, test_split_loose_owner, test_spline_thin 12/12, test_part_list 41, test_pl_select 28, test_split_side_views 15/15. No pladd/plnew tests existed (scripts /tmp/o91/). F1 n/a (operator door; heal.py untouched). Ramifications: only these two doors; heal/upload/nest/post untouched. Found, NOT fixed: a row merged as identical copies (qty 17) brings its picked copies in once -- pre-existing, separate. -- 2026-11-04 22:45 UTC Jordan's test on a fresh grill (list grill-with-bend-lines-original-1793831946): Add to row of all items to row 1 was fast and complete, but the 64 emptied rows stay on the list (one-by-one delete) -- asked: auto-offer 'remove empty rows' vs bulk row delete. Undo then New part from selection (move, 973 items) -> part-66-1793832243: heal flagged (5 jogs, fidelity 0.0049), QC 132 flags (58 branch, 33 dangling, 27 duplicate, 21 open from doubled ink, 14 gap). Jordan: 'making this file usable is still a daunting task'. Likely (not yet measured per flag) bend lines picked in + lines drawn twice (step 4). -- 2026-11-04 23:52 UTC note for step 3: a Cancel already exists (CANCEL_DOORS: pack, plan, post, sheet post, edit save, tolerance, part scale, units, upload); Add to row / New part / box pick are NOT among them (they write mid-run). Step 3 = join them to the existing Cancel, not a new feature. -- 2026-11-05 SHIPPED to main d21b79c9 by copying dev's server.py/app.js/test_sheet_window.py (= main + exactly O90/O91/O93, checked by patch replay; O89 heal.py not shipped). -- 2026-11-05 00:13 UTC BUILD (O94): Add to row (pladd) and New part from selection (plnew) join the global Cancel (CANCEL_DOORS + client CANCEL_DOORS 'Cancel adding' / 'Cancel the new part'). The doors poll the cancel before their first write and pass it into every run_pipeline; on emit.Cancelled the handler calls the new _undo_rollback(): every file the door touched goes back to its before bytes (the same snapshot one Undo uses), a job directory the door created goes away, nothing is journalled. Tested on a scratch engine (:8491, temp copy of grill list 1793813710): cancel at 4 s and at 15 s, both doors -> 0 files differ (whole job tree, byte hashes), undo journal unchanged, no new job dir; a run without cancel still adds (24.2 s, 265 contours) and journals one 'Add to part' step; browser: busy card shows 'Cancel the new part', click -> toast, list unchanged. Identity gate green; test_split_door 17/17, test_part_list 41, test_pl_select 28, test_plan_cancel, test_pl_cancel 11. test_upload_cancel 11/15 on dev -- proven NOT O94 (dev without O94 also 11/15; O94 with main heal.py 15/15): it is O89 (see that pin). F1 n/a (operator doors).