first commit

2026-07-17 17:45:56 +08:00
commit 04c487823b
5873 changed files with 12227228 additions and 0 deletions
@@ -0,0 +1,93 @@
+# BrepFaultProbe Lessons
+
+This file records the failure patterns that have already appeared in this project.
+
+## 1. Do not require explicit labels when the geometry can be inferred
+
+The sleeve-removal `b` case failed at first because the judge only accepted edges explicitly tagged as `push_face`.
+
+Correct rule:
+
+- accept explicit labels when present
+- also infer the same feature from B-rep structure
+- treat explicit labels as strong evidence, not the only evidence
+
+## 2. Keep candidates separate from decisions
+
+`*_candidate` is evidence, not a final answer.
+
+If the pipeline mixes candidate evidence with final template matching, the judge will drift into false negatives or false positives.
+
+## 3. Outer frame is not a functional feature
+
+For 2D sketch graphs, the outer frame can be a valid drawable loop, but it does not by itself mean the image contains a meaningful machine-design feature.
+
+Always separate:
+
+- background outer frame
+- actual functional loops
+- inferred feature loops such as holes, slots, bosses, and ports
+
+## 4. Image and model must share a canonical middle form
+
+Raw raster edges and true section edges are not directly comparable.
+
+The stable comparison point is:
+
+- vertices
+- drawable edges
+- paths
+- loops
+- regions
+- relations
+- dimensions
+
+## 5. Hatch lines are annotations
+
+Section hatches must be excluded from drawable contour edges.
+
+If they are treated as normal edges, the sketch graph becomes noisy and the loop/path matcher starts inventing structure.
+
+## 6. One sample is not enough
+
+A rule that works on one positive case may still fail on:
+
+- unrelated cases
+- weak candidates
+- cases with no explicit role tags
+- cases whose functional feature is expressed by a different geometric cue
+
+Always verify at least:
+
+- a positive case
+- a negative case
+- an unrelated case
+- a weak/noisy case
+
+## 7. Current open risk
+
+`b` has been upgraded to infer push-face evidence from B-rep topology.
+
+`c` and `d` still depend more on explicit region semantics, so they can repeat the same failure mode if the upstream extractor does not label them clearly enough.
+
+## 8. Textbook image rules must name the exact measurable geometry
+
+Failure pattern:
+
+- A textbook rule says a "diameter" should be larger, but the drawing contains many diameters.
+- If the extracted knowledge only says `boss_diameter > cover_diameter`, the B-rep pipeline may measure the wrong diameter, such as an inner hole, register-fit diameter, bearing-seat bore, bolt hole, or bolt-circle diameter.
+
+Correct rule:
+
+- The knowledge chunk must explicitly define the compared geometry by position and boundary role.
+- Use terms such as `mounting-side outer contour`, `maximum outer diameter`, `cast boss outer boundary`, and `minimum boundary margin`.
+- Also list excluded diameters when they are plausible distractors.
+
+For Chapter 18.2.1, the rule is:
+
+- compare the end cover mounting-side flange outer boundary or maximum outer diameter
+- against the cast housing boss/recess/allowance-region outer boundary or maximum outer diameter
+- require the housing cast boss boundary to contain the cover flange boundary with casting-position-error allowance
+- do not compare the cover inner hole, register-fit diameter, bearing-seat bore, bolt holes, or bolt-circle diameter
+
+Every image-to-rule extraction should preserve this level of specificity so model-side B-rep facts can be selected without ambiguity.