Native move-entry memory comparison and bounded frame capture
============================================================

English baseline12 and final14 run the supplied native1.2 runtime with the same
memsize=1 MiB, conventional-memory/disk-only input profile, DOSBox Staging 386-class normal core and read-only
UI4-derived handle observer. This is logical GEOS handle accounting, not physical
8088 RAM, 8086 instruction acceptance or a full gameplay peak measurement.

report.json and the top-level native-control-memory-comparison.json are identical.
Identities and snapshots bind both executable hashes, observer source/library
hashes, emulator hash and normalized input profiles. Raw workspace probe CSV hashes
are recorded; large continuous CSV files and the user runtime are not shipped.
Snapshots retain one-second sample windows, medians and ranges for every stage.
Before loading, system resident bytes differ by -32 (340640 vs 340672), retained
as observed runtime/cache variability; it is not silently normalized away.

At idle app-owned allocated/resident bytes increase by 832; HF_FIXED increases by
240. The largest sampled stage window is 219936 app-owned bytes (+1424 vs baseline
idle). Twenty type-five/select-all/delete cycles end at 219328 bytes, stable across
the last ten cycles, with no increase in HF_FIXED or handle count. This bounded
check does not establish every editor history or all possible leak behavior.
Other-owned and system resident changes include shared library/cache/GeoManager
state and cannot be attributed purely to the field. Residency varies with paging.

Only after all memory snapshots, a passive host SDL texture-presentation observer
captured ordinary Enter and Play e2e4 transitions. All 448 presented samples preserve
every pixel in the 60 unchanged squares. The 6 unique frames are lossless PNGs;
frames.csv records every presentation and time. No sampled frame shows an all-white
board. This is bounded observation, not universal no-flicker proof.

reference-tools contains exact source/scripts with original separated harness
workspace paths. Reproduction needs one's own compatible runtime and emulator/Xvfb;
no runtime, font, emulator or host binary is redistributed. The figure generator
needs matplotlib; frame analysis uses Pillow. Report/snapshot analysis uses the
Python standard library. No guest RAM patch is made.
