8 GB, DM_GUEST_FILE=0: 8.3 minutes to the 4800-4950 MB crash band under heavy scene churn, a ~2 minute playable window inside it, and which baseline should we re-test on #2575

New data since #2319 (new issue rather than a comment, as preferred). Environment unchanged: 8 GB M2 MacBook Air, Fire Emblem Fortune's Weave (04007EF01F8AE000), DM_GUEST_FILE=0, evict LRU on (IDLE=600/EVERY=120/AT=3500/KEEP=16/CAPMB=512).

1. The crash band is reached in 8.3 minutes when the route is heavy. Same save, same city, but continuous running and frequent scene changes:

• 11:39:26 start -> 11:42:35 first warning, residency 3866 MB (+793 MB in one step) -> 11:45:43 second warning 4564 MB -> 11:47:47 residency 4981 MB, FATAL GPU out of memory, 32 command buffers failed, exit 134.
• Rate 395-483 MB/min against 130 MB/min on the calm 86-minute session that peaked at 4105 MB. The ceiling itself did not move; only the time to reach it did, and it tracks how many new scene resources get registered per minute.

2. Inside the band there are only about two minutes of playable time. Between the two warnings the frame time degraded, and in the last 527 s the long frames were 42.6 / 65.7 / 83.8 / 92.5 ms with the GPU steady at 26-38 ms and submit+present at 10-53 ms, i.e. the host side was waiting; memstat showed a footprint around 10.1 GB against 8 GB of RAM. So after the first warning a player has roughly two minutes before the process dies, which is enough to save but not to keep playing.

3. What we would like to know.

• Which baseline should we re-test on? We are 293 commits behind and our shader chain (b-1d619d0f520b) is no longer in upstream history. You have recently pointed at 529cf68+ in Ocarina of Time on an M1 Pro: slow motion after ~80 min on main 25c2546, finalized-texture table puts jump from <1 to thousands a frame #2213, chain a5e7fb6e4fa4 with 034e91a + 596d719 in Ocarina of Time (M1 Max): fog, light shafts and glow are no longer drawn between d8b2b64a and ced36121 #2276, and PR Ocarina of Time draws its 3D above its Switch 2 resolution: Render scale 1.333 gives native 2560x1440 and 2.0 native 3840x2160 docked, where the console draws 1920x1080 and upscales #2532 / DM_NVN_SHADER_ALT: an in-run A/B of any translated shader variant, and fsbench, an offline benchmark of one pixel shader #2315. Rebuilding our shader library takes hours, so we would rather do it once: is there a recommended combination for a non-Ocarina title on an 8 GB machine, or are those baselines specific to the games they were measured on?
• Is Degrade instead of abort when the GPU allocation budget runs out (opt-in), instead of a bare exit 134 #2522 (degrade instead of abort when the allocation budget runs out) still something you would consider? If it is, we can test a variant on the same 8 GB machine that produced these numbers.

Logs and memstat for the session are available on request.