diff --git a/.claude/memory/MEMORY.md b/.claude/memory/MEMORY.md
index 0361056..9a8a793 100644
--- a/.claude/memory/MEMORY.md
+++ b/.claude/memory/MEMORY.md
@@ -35,6 +35,7 @@
 - [Boot splash and logo](schunix-boot-splash-logo.md) — the loader draws the logo, not the kernel; STAGE2_SECTORS was in five places and is now in one.
 - [SchAPI direction](schunix-schapi-direction.md) — Schunix as a macOS-shaped *system*; the stable ABI is a library, not the syscall table.
 - [Schcom v0](schcom-v0.md) - the compositor runs; Wayland-shaped, verified by checking pixels from the host.
+- [SchFS next + platform roadmap](schfs-next-and-platform-roadmap.md) - snapshots/scrub/compression/encryption, one home for the format, a unified loader, audio; and the 614 MB ceiling that was fixed.
 - [Roadmap and projects/](schunix-roadmap-and-projects.md) — the agreed order through to Schcom/SchWM/SchDE; the GUI and installer live outside the OS tree.
 - [Driver infrastructure](schunix-driver-infrastructure.md) — dynamic IRQ registration, a DMA arena, PCI capabilities; PCI interrupts are level/active-low.
 - [Network stack choice](schunix-network-stack-choice.md) — don't write TCP; lwIP over FreeBSD over NetBSD, with the licence evidence.
diff --git a/.claude/memory/schfs-next-and-platform-roadmap.md b/.claude/memory/schfs-next-and-platform-roadmap.md
new file mode 100644
index 0000000..7150523
--- /dev/null
+++ b/.claude/memory/schfs-next-and-platform-roadmap.md
@@ -0,0 +1,67 @@
+---
+name: schfs-next-and-platform-roadmap
+description: "The 2026-08-10 direction — SchFS features, one shared home for the format, a unified loader, audio; and the two requests that were pushed back on"
+metadata: 
+  node_type: memory
+  type: project
+  originSessionId: d4821aa9-2e49-40da-91e7-90ef9ed9997e
+  modified: 2026-08-10T16:52:05.449Z
+---
+
+**Set out by Nexoniarz 2026-08-10**, after the compositor and resolution work.
+
+## Two requests that were pushed back on, and why
+
+**"Make SchFS bigger than ZFS."** Declined as stated, and Nexoniarz did not
+press it. SchFS already addresses **64 ZB** (64-bit block numbers at 4 KB).
+ZFS is 128-bit — a number its designers said could not be filled without
+boiling the oceans. Beating it means 128-bit block pointers: every pointer in
+the B-tree doubles, cache pressure with it, for a limit already physically
+unreachable. A spec-sheet win with a real cost.
+
+**"Move SchFS to projects/."** Cannot be done as stated: `projects/` is for
+things that run *on* SchOS, never linking kernel headers and never a dependency
+of the kernel build ([[schunix-roadmap-and-projects]]). SchFS is the root
+filesystem — the kernel must mount it to boot. **The version that works** is a
+neutral `lib/schfs/` consumed by the kernel, by `mkschfs`/`schfs_fsck`, and by
+`projects/schfs-linux`. The source is *already* shared verbatim; this only moves
+where it lives. Agreed as the sensible form, not yet done.
+
+## Done 2026-08-10: the ceiling that actually existed
+
+The format said 64 ZB; the stack under it said **614 MB**. `idestart` panicked
+on any block past `FSSIZE` (150000 blocks) — a constant describing a v1
+filesystem *image*, bounding a physical device. Fixed in commit 527e6ea along
+with the rest of the chain: `struct buf.blockno` and `bread` are 64-bit,
+sectors are computed in 64 bits, and LBA48 is used per request when the sector
+does not fit in 28 bits.
+
+**Still capped, deliberately:** raw `/dev/hdN` goes through devsw's 32-bit
+offset, so a raw device is addressable to **4 GB**; `usb_storage` builds a
+32-bit LBA, which is what SCSI READ(10) carries — READ(16) is a different
+command, not a wider variable.
+
+## The SchFS feature list, all four wanted
+
+Nexoniarz asked for all of these. Suggested order is mine, cheapest and
+highest-leverage first:
+
+1. **Snapshots and clones.** The copy-on-write B-tree already makes a snapshot
+   a retained tree root ([[schfs-btree-phase1]]). Almost certainly the best
+   value per unit of work on the list.
+2. **Integrity: scrub, repair, redundancy.** CRC32c per data block already
+   exists; this is detecting and *repairing* rather than only reporting —
+   duplicate metadata, a scrub pass.
+3. **Compression**, per extent. The extent record already carries a length and
+   a disk block, so there is somewhere to put a compressed size.
+4. **Encryption.** Confidentiality, where the checksums give integrity. Last
+   because it is the one with the most ways to be subtly wrong.
+
+## The platform list
+
+- **A unified second stage.** BIOS and UEFI become thin starters handing off to
+  one common loader. This is worth real money: the duplication is exactly what
+  produced the `EFER.SCE` and RSDP bugs, where the BIOS path had something the
+  UEFI path silently did not ([[schos-uefi-loader]]).
+- **Audio.** First sound driver, new device class. Intel HDA or AC'97 — both
+  emulated by QEMU, both with public specs.