[jpudysz]

Problem

collectVisibleAncestors and collectAncestors kept their per-call caches in plain arrays indexed by stateNode.canonical.nativeTag. Native tags only grow: every view ever mounted takes a new one. Each call writes hundreds of scattered, large tag indexes into a fresh array. As tags grow, Hermes allocates far more for those writes than for a Map/Set, so the cost grows with session length.

Both consumers call these after every theme change:

• Unistyles: collectVisibleAncestors in the ShadowRegistry refresh.
• Uniwind Pro: collectAncestors in refreshReactNodes.

The garbage shows up as extra GCs on unrelated work, like re-renders.

	Before (array)	After (Map/Set)	Paired
Fresh: JS allocated	525 MB [508-535]	472 MB [454-487]	-10.3%, 12/12
Fresh: GC count	148 [143-151]	133 [128-139]	-10.5%, 12/12
Fresh: GC time	188 ms [162-250]	177 ms [147-209]	-8.4%, 10/12
Long: JS allocated	1431 MB [1324-1467]	475 MB [441-489]	-66.6%, 12/12
Long: GC count	222 [201-227]	139 [130-146]	-36.9%, 12/12
Long: GC time	381 ms [357-564]	202 ms [155-279]	-47.7%, 12/12
Long: GC time during mount + re-render	219 ms	158 ms	-30%, 12/12

One collectAncestors call over 900 hosts

Native tag reached	Before	After
~2k	44 kB, 0.19 ms	62 kB, 0.21 ms
~50k	419 kB	62 kB
~550k	4.0-8.3 MB, 0.69 ms	62 kB, 0.21 ms

Per flip

Before, flip allocation grows every round:

• Fresh session: 1.67 MB to 2.83 MB (flip to dark), 3.3 MB to 5.7 MB (flip to light).
• Long session: about 9.9 MB / 19.8 MB.

After, it stays flat at 1.68 MB / 3.35 MB.

Wall-clock

Mount (~25 ms), re-render (~19 ms) and unmount (~4 ms) show no difference beyond noise. A flip window (40-90 ms, setTheme to the next idle callback) is dominated by native work and idle scheduling, so it can't resolve a sub-millisecond JS change. With fresh tags the Map/Set version costs about +0.02 ms per call.

Earlier measurement in Unistyles (900 views, 28 rounds per launch, A/B against v3.5.3): total allocation went from 2181 MB to 1952 MB with this change, and GC time dropped about 8.5% below v3.5.3.