Improve L2 market parsing and data handling

2026-08-11 22:24:36 +03:00
parent de41230b9b
commit af9cd1e92d
58 changed files with 2474 additions and 444 deletions
@@ -0,0 +1,39 @@
+# Dynamic Tooltip Crop Evidence
+
+## Acceptance Evidence
+
+- Browser geometry: a 1920 x 1080 reference keeps an intrinsic canvas of
+  1920 x 1080 and renders at 3840 x 2160. The stage scroll width grows to
+  3888 px, so the image is not squeezed back into the panel. See
+  [calibration-2x.png](./calibration-2x.png).
+- Browser pointer smoke: drawing on the 2x canvas creates calibration regions
+  and advances through all six calibration steps. The marker drag from display
+  `(326.6, 604)` to `(426.6, 642)` mapped to source `(288, 259)` through
+  `(338, 278)` and was subsequently found by anchor matching at 99.8%.
+- Real tooltip fixture: the complete calibration check finished 6/6 stages and
+  parsed `Sword of Revolution`, quantity `1`, and `1,700,000 Adena` at 70% OCR
+  confidence. The third descriptive tooltip row was not passed to item parsing.
+  See [tooltip-diagnostics.png](./tooltip-diagnostics.png).
+- The wide tooltip region is used only for the `Price :` anchor search. The
+  item OCR preview contains the fitted name and price rows, not the full search
+  band.
+
+## Verification
+
+- `node.exe --test`: 18 tests passed, including white/yellow pixel masks,
+  stopping a text run before distant UI noise, and fitting a short row whose
+  legacy saved width crosses the frame edge.
+- `node.exe node_modules/vite/bin/vite.js build`: production build passed.
+  Vite reports the existing large OpenCV chunk warning.
+- `git diff --check`: passed; Git only reported LF-to-CRLF conversion warnings.
+
+## Review Notes
+
+- PRE plan review: aligned after clarifying the single owner of the dynamic
+  right edge, the no-wide-OCR fallback, and legacy calibration behavior.
+- POST plan review: aligned; saved screenshots and the unbounded-run regression
+  close the evidence and no-wide-fallback gates.
+- Correctness review: no remaining findings after isolating fitted rows in a
+  masked composite and applying yellow pixels only to the price row.
+- Maintainability review: no remaining findings after reading pixels directly
+  from the source canvas and simplifying the text-run control flow.