Docs: G15 and the parity doc cite budget runs and slowness reports since 2026-10-11

The "Performance unverified" bullet in docs/target-app-parity.md said
that the budgets had not been run on an idle machine since 2026-10-01,
that the last loaded run missed the 16 ms frame budget at 50k paths,
and that no user had reported slowness in mainstream work beyond #575
and #758. On 2026-10-11 #1155 recorded idle-machine runs, made before
#1173 merged, and runs of #1173's change met every budget at 50k paths
(load average 4 to 7) and at 25k paths. #1097 (zoom and pan in 0.9.0)
and #1731 (a drag that stutters as it starts) are open. G15 in
docs/gaps.md called the 25k-path run the last one, but on 2026-10-13 a
comment on #1172 reported runs of 0.10.0 on Linux with 4 logical CPUs
that missed all three render budgets at 20k paths.

The bullet now gives the 25k-path run and the runs in #1172, and lists
#1097 and #1731 with #575 and #758 as open user reports of slowness.
G15 no longer calls the 25k-path run the last one and adds the runs in
#1172.

docs/gaps.md

@@ -218,8 +218,9 @@ from G14 and the budget run of G15: **~170–270 h** one agent, ~50–80 h wall
 ## G15 Performance budgets and a GPU renderer
 
 - **Improved (#1257):** Craquelure, Mosaic Tiles and Stained Glass reuse squared distances during nearest-cell searches, with the original distance ordering retained near rounding/tie boundaries. A brute-force boundary regression and full-viewport pixel comparisons agree with the old result. `texture_drag_perf` reproduces a moved CMYK rounded rectangle at 190.07% and DPI 1.5; timings from a busy development machine are supporting evidence, with idle Windows drag QA still pending.
-- **Missing:** an idle-machine run of `vectorcraft-cli perf` at 50k paths (last: 2026-10-11 at 25k paths, every
-  budget met, 400% zoom 12.3 ms against 16 ms, measured), a spatial index for zoomed-in frames, dirty-region
+- **Missing:** an idle-machine run of `vectorcraft-cli perf` at 50k paths (2026-10-11 at 25k paths, every budget
+  met, 400% zoom 12.3 ms against 16 ms, measured; runs of 0.10.0 on Linux with 4 logical CPUs, reported in #1172
+  on 2026-10-13, missed all three render budgets at 20k paths), a spatial index for zoomed-in frames, dirty-region
   rendering, the GPU backend spike (M5). Large-selection edits, cached live effects and Pathfinder Trim, Merge,
   Crop and Outline landed in #1173.
 - **Not redrawn:** the canvas keeps drawing these objects as they were:

docs/target-app-parity.md

@@ -168,9 +168,11 @@ The discounts, with the evidence:
 - **Stability on real machines (×0.92).** The code never panics by design, and the main platform (macOS) starts
   reliably, but 26 issues mention a crash, freeze or failed launch (8 open: #858, #964 fixed the day of this
   review, #620, #818, #905 on Windows 7, #834). There is no field record and no real-world file corpus yet.
-- **Performance unverified.** Budgets (`vectorcraft-cli perf`) haven't been run on an idle machine since
-  2026-10-01; the last loaded run missed the 16 ms frame budget at 50k paths. Not discounted separately: no
-  user has reported slowness in mainstream work beyond #575 (Linux AppImage) and #758 (`.ai` open time).
+- **Performance unverified.** Budgets (`vectorcraft-cli perf`) were all met at 25k paths on 2026-10-11; runs of
+  0.10.0 on Linux with 4 logical CPUs, reported in #1172 on 2026-10-13, missed all three render budgets at 20k
+  paths. Performance is not discounted separately. Open user reports of slowness in mainstream work include #575
+  (Linux AppImage), #758 (`.ai` open time), #1097 (zoom and pan in 0.9.0) and #1731 (a drag that stutters as it
+  starts).
 - **Missing subsystems** a minority need daily (3D, the Effect Gallery, CJK vertical composition, scripting)
   pull the full number below the mainstream one.