[rash-pro]

Fixes the washed-out Deku Tree exit after Gohma from #508 (closed as a duplicate, tracked in #1004): leaving through the blue warp, the forest came out washed white, Link and the plants glowing, until the Deku Tree's cutscene. Still there on v2.0.11.

Cause

mirror_make fills a remade pool's new mirror from guest memory, then copies each retired pool's mirror over its whole part. This is the inheritance from e900faa: what the GPU wrote there and nowhere else, the foliage records. That whole-range copy also puts the old mirror's words back over pages the CPU wrote since that mirror's last sync. The captures that go over it afterwards hold only the last submissions' pages, while a scene change takes many. Through the blue warp, the dungeon's values came back over what the game wrote while the next area loaded.

Change

• The CPU's writes get a page generation of their own (stamp_cpu_page), stamped where memory_collect_written stamps the collected pages and where a capture stamps its pages. memory_stamp does not stamp it, because the GPU's own writes use memory_stamp.
• The inheritance copies the runs of pages the CPU has not written since the retired mirror's synced_gen, skipping clean 2 MB regions whole.
• What the GPU wrote is inherited as before. A page the CPU wrote stays guest memory's (or a capture's), as the old mirror's next sync would have made it.
• RX_VK_INHERIT_WHOLE=1: the whole range, as before.
• tests/test_inherit_cpu_pages.py checks the two stamps, that memory_stamp does not take the CPU's generation, and the inheritance's test.

Using the plain page_gen instead also fixed the haze. But it counts the GPU's own stamps: it left thousands of pages per inheritance un-inherited, against 0 to 267 with the CPU-only generation. That would put the foliage records at risk again.

Testing

RTX 4070 Laptop (8 GB), NVIDIA 615.78.08, Linux (Wayland). The change was tested on top of v2.0.11, with only hookshot_vk.so rebuilt (RX_VK_ONLY=1). On main it builds, and the new test and the existing checks that read vk_memory.cpp pass (tests/run_all.py).

The setup:

• An autosave standing at the blue warp, run headless.
• HOS_INPUT_SCRIPT presses Continue and Load, walks into the warp and clears Navi's line.
• RX_VK_SHOTS writes a shot every 20 presents.
• The score is the median luma of the shots in the 2 s after the warp's white fade.

Build / switch	Median luma after the warp
v2.0.11 (3 runs)	197, 203, 197 (washed white)
v2.0.11 + this (2 runs)	85, 86
v2.0.11 + this, RX_VK_INHERIT_WHOLE=1	203
RX_VK_NO_INHERIT=1 (2 runs)	87, 89
RX_VK_MIRROR=0	94, 101 (no haze, but Link and the plants drawn as white silhouettes)
RX_VK_WB_DIFF=0, RX_VK_MIRROR_MB=1024, RX_VK_NO_READBACK_WAIT=1, RX_VK_WAIT_GPU=1, RX_VK_MIRROR_KEEP=1	186-205 (unchanged)
v2.0.11 + 9928d74 (BC targets through raw views)	204 (unchanged)

Also played through the warp with the fixed build: the forest is lit normally and the Deku Tree's cutscene starts as usual. Not checked: a wide Kokiri shot of the tree canopies, which the inheritance exists for. The change keeps everything the GPU wrote, so they should be as before.

Draft until an interleaved fps A/B is in. The headless runs above drifted with temperature over a long series, by up to 4 fps between runs of the same build, so they are not a fair before/after for fps.

🤖 Generated with Claude Code

[rash-pro]

Closing: #1127 landed the same fix while this was in draft (same root cause: the inheritance's whole-range copy over pages the CPU wrote since the retired mirror's sync), and it also covers the parallel-replay case this one missed.

I checked #1127 against the case this PR was for, #508's washed-out Deku Tree exit after Gohma. I cherry-picked 0266268 onto v2.0.11 (RTX 4070 Laptop, NVIDIA 615.78.08, Linux) and ran the same headless warp test as above. Median luma in the 2 s after the warp's white fade:

Build	Runs	Median luma
v2.0.11	3	~200 (washed white)
v2.0.11 + 0266268	2	90 and 87 (lit normally)

So #1127 fixes #508's case as well.

One difference, in case it is useful. This PR stamped the CPU's writes with a generation of their own, so pages stamped only by the GPU's own writes kept being inherited. An example is the first page of a bare table entry: its first 64 bytes are marked written, but the rest of the page holds records that are never written back. #1127 tests page_gen, which those stamps raise too. That matches what mirror_sync already does, so it is consistent, but it can take guest memory for such a page. I did not see any difference in play from it.