[Jozkah]

The Reverse engineer items from #34, rebased on 0.9.3. The notes fix is left out because you shipped it in 0.9.3. One commit per item, each with a regression test that fails without it:

• A project with an Import STEP feature never read as saved. Each capture encoded the STEP texts into new bytes, so the snapshot signature never matched. The leave-page warning always fired, project.unsaved was always true, and recovery rewrote every member every 2 s. The bytes are now reused while the texts are the same.
• Duplicate feature ids from a project file. nextFeatureId / nextParameterId were only checked to be finite, so a file could set them at or below an id in use, and removeFeature then deleted both. They're moved past the largest id on load, and a non-integer nextOfKind is refused.
• STEP file lookup by inherited keys. An Import keyed constructor or __proto__ found an Object member. Lookups now use Object.hasOwn.
• The kernel kept the pre-edit scan. After a Mesh Tools copy was adopted, or undo crossed it, the file name and triangle count didn't change, so syncScan didn't send the new scan, and splits and compares ran on the old one. A scan-change counter now takes part.
• Build counters drifting. A throw in finishBuild counted a build off twice, and readRegionsAgain had no catch. Each build is now counted off once.
• A failed helper kernel load was cached. The helper now tries again on the next request.
• Deep expressions and long parameter chains overflowed the stack. The parser has a depth limit, and expressions over 4000 characters are refused. Parameters are read in dependency order, and an overflow becomes an expression error.

Build: npm run build passes. npm test (3104), npm --prefix mcp test, the plugin boundary and units checks, and npm run e2e:ci all pass.