[wieslawsoltes]

Repeated same-class mutations in unchanged Code OSS were still feeding every matching final-compound selector through the routed invalidation planner. A typical startup transition scanned 800–1,300 dependency entries even when the changed element was the only possible invalidation subject.

This change sends exact final-compound subject dependencies directly to the subject cascade queue while retaining the compiled planner for ancestor, descendant, child, sibling, and fallback routes. It also adds an A/B escape hatch, bounded tracing, a 64-subject/512-rule scaling regression, and a dedicated CTest lane.

Advances #243.

Result

On unchanged Code OSS 645f29cc, exact WebScene a99eaf2, and AppScene cd0a02e:

• Workspace paint gate: 9/20 to 20/20 under 2 seconds
• Resolve p50: 914.5 ms to 751.6 ms (17.8% faster)
• Resolve p95: 1158.5 ms to 762.6 ms (34.2% faster)
• Paint p50: 2164.3 ms to 1828.2 ms (15.5% faster)
• Paint p95: 2262.0 ms to 1860.1 ms (17.8% faster)
• Exact Explorer model/rows/DOM binding: 20/20

The product-neutral regression records 65,664 planner lookups on the old path and zero on the direct path, with the same 128 cascades and computed styles.

Validation

• webscene_native_direct_subject_invalidation: passed
• attribute-invalidation-scope: passed
• nested-functional-selectors: passed
• relative-sibling-has: passed
• AppScene SDK producer tests: 31/31 passed
• macOS compiler/runtime profile: passed
• Release app codesign verification: passed
• Unchanged Code OSS exact product gate: 20/20 passed

The broad existing filters still contain unrelated baseline failures: live-form-state counts both the invalid form and control, and broader invalidation scaling has existing memory/tree-reaction ceilings. This PR does not change those paths.

Detailed evidence: docs/validation/direct-subject-selector-invalidation-243.md