[Japabu]

FS phase 2, stage 3a, stacked on #1141, whose write-span version it publishes. The rulings are the orchestrator's: a per-holder stat syscall; the kernel owns size and version; the owner writes mtime as a word on Write; one counter. Review only the top commit until #1141 lands.

What changes

• SYS_PAGED_STAT(holder) -> PagedStat {size, mtime, version}, read without a request. The size and version are the kernel's own. It is atomic against the pager's writes, as a read is: while a write is under way it answers WouldBlock, and the SDK asks the server instead.
• The SDK keeps the open reply's stat beside the holder. A held descriptor's fstat is the object's size and mtime over that reply's kind, node, volume, nlink, atime and ctime. No change to the file moves those.
• The owner writes mtime. It travels as PagedArgs.time on Write and Resize, as an argument to paged_create, and through a new PagedOp::Stamp for a utimens.
  • The file server hands over the volume's own value after every change that moves it: a reserve, a truncate, an open that truncates, a set time, and a refused write's cut-back. So the precision is the volume's.
  • Stamp is the one op beyond the ruling. A utimens moves no bytes or size, and no other call carries a time.
• FAT's mtime is now its last write's. It used to be the entry's last level, so through a handle it disagreed with the same file reopened (issues/a-fat-files-mtime-through-its-handle-is-its-last-level.md, closed). That made it unpublishable. A write or resize now records its even second, which the handle answers and the level stamps.

Commits

Rebased onto #1141's r2 head 029ca221c (main 058bea2). SYS_PAGED_STAT is now 164, numbered against main (whose last is SYS_CURRENT_CPU = 163); SYSCALL_PROFILE_BINS is 166.

• ea23ee415 stage 3a: kernel, ABI, SDK, file server, the FAT mtime fix and a guest member.
• ca6340ef3 issues: FS stage 3b waits for memory stage 7b's mappings (records only).

Checks

Evidence after the rebase: https://github.com/ToyOSOrg/evidence/tree/f589db58a6b52c475fa6bac957352ffb046239c1/r2 (r0: the original checks and controls; r1: the metal A/B).

• cargo run -- --ci host at ca6340e: rc=0, 109 steps green.
• QEMU x86_64, the testcases boot via --debug, at ca6340e with a clean tree: rc=0. The new a_held_files_fstat_is_its_servers_stat exits 0, as do the tear member, a_served_file_reads_through_its_paged_object, served_file_offset, a_pager_s_end_answers_its_waiting_holders_gone, std_fs, file_mtime and stat_identity.
• Why the new guest test. A held fstat and the server's stat come from the kernel's object and the file server's volume. Only a guest runs both. Each side's rule is host-tested. Targets: QEMU x86_64 (run), T14 and AMD (staged).
• Negative controls, all built and red:
  • host C8–C12 (the server's mtime hand-over, FAT's write time, the pure model's publication);
  • guest G2 (the kernel publishes mtime 0), which also shows the held fstat is the kernel's answer.
• Metal: run before the rebase, claim met (r1). fs.disk.fstat went from 9.63/8.17 to 0.361/0.361 us on AMD and from 10.11/10.34 to 0.452/0.452 us on the T14 (−96%). The rebase moved only the syscall number, so that result stands. The rule-2 flags (AMD write_4k_fsync +8.3%; T14 create_write_unlink and mkdir_rmdir, disk-mutation rows 3a's code doesn't touch) match the boot modes paged: a holder's read is atomic against its pager's write; an fsync commits its own file #1141's r2 documents. Every testcases member passed on both benches.
• Controls C8–C12 and G2 (r0) ran on the pre-rebase tree. The code they mutate is unchanged.

Risks

• An mtime is published only when the file server hands it over. A future path that moves a volume's mtime without calling Files::stamped, write, truncate or truncated would leave holders stale. The host test covers every path that exists today.
• A stat during a write costs a round trip. It falls back while any write to the file is under way.

🤖 Generated with Claude Code

[Japabu]

Superseded by #1152, the storage consolidation. This PR will be closed when #1152 lands.

Its last pushed head, ca6340e, numbered SYS_PAGED_STAT 164. Main has since taken 164 (SYS_DEVICE_FIRMWARE_WINDOW), so whatever carries this forward needs the next free number. Its metal claim (fs.disk.fstat −96% on both benches) is in https://github.com/ToyOSOrg/evidence/tree/50415279a8a7b88a9f019adeeb2dea8e0df9e16c/r1, and the checks after the last rebase are in https://github.com/ToyOSOrg/evidence/tree/f589db58a6b52c475fa6bac957352ffb046239c1/r2.