Optimize eligible sketch-instance execution and cancel native work

[ksbharaj]

Summary

Render eligible first sketch instances without executing unrelated solids and booleans, and stop background native execution when the Python await is cancelled.

• Add try_render_sketch_instance and its code-string counterpart. Unsupported cases return None so MCP can use ordinary execution.
• Reuse the existing stop-after-sketch-solve mechanism while retaining function arguments, normal unit/constraint evaluation, source locations, and the selected instance.
• Keep the existing conservative eligibility scope together: direct first-instance calls, literal inputs/arithmetic on earlier literals, supported sketch operations including sin, and an optional named region feeding the helper's extrusion.
• Reuse the shared Python/native cancellation implementation now merged in Cancel native KCL tasks with Python callers #13859, retaining in-flight cleanup and subsequent-call regression coverage. There is no new native timer: MCP owns the call's single 120-second budget.

Scope and dependencies

Split from #13734. This PR is stacked on #13882 (codex/sketch-instance-selection); its diff is only performance and cancellation regression coverage. The point-colour change in #13734 is independent. No region/segment visualization is included, and this split adds no new optimization cases or behavioural changes. A subsequent safety fix declines the shortcut if removing a referenced outer binding could change the sketch’s meaning, falling back to ordinary execution.

The previously tested combined commit 70b856e29374d84ca2f0ca5707e0b2b427b551c9 is preserved on codex/sketch-visualization-tested-combined. Existing MCP/TTC pins stay there until the reviewed pieces are available together in a published native release.

Reviewer focus

• Correct instance, not just a faster image: preserve geometry, constraints and placement. Imports, aliases, face-dependent sketches, later indices and unsupported helper layouts keep ordinary execution. This is not a general dependency slicer or cache.
• Shared cancellation behaviour: uses main's spawn_py implementation from Cancel native KCL tasks with Python callers #13859; this PR no longer changes that wrapper. Cancellation is cooperative. In-flight integration tests relay a real DEV engine connection, confirm socket closure after direct cancellation or timeout, then check a subsequent call succeeds.
• Bounded scope: the helper-layout and operation allowlist are deliberate restrictions. A new unsupported spelling should not silently render a different sketch.

The equivalence regressions compare full and optimized PNG bytes, constraint reports and placement. Saved real Flanges checkpoints motivated the optimization: two historical 519/619-second calls rendered byte-identical PNGs in approximately 1.7/1.8 seconds with the shortcut. These are focused tool benchmarks, not a controlled claim about total Zookeeper runtime. On the same newer build, a full-execution comparison hit its 240-second cap, so no precise end-to-end speedup ratio is claimed.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
modeling-app	Ready	Preview	Sep 15, 2026 5:00am UTC

[codspeed]

Merging this PR will not alter performance

✅ 150 untouched benchmarks
⏩ 128 skipped benchmarks¹

---

_{Comparing codex/sketch-execution-performance (b91eaeb) with codex/sketch-instance-selection (5276a3f)}

Footnotes

1. 128 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩