[AlexanderNyr]

ioc: power_on clears only the lines the IOC drives

Ioc::power_on stored 0 over the CPU's whole interrupt word. That word is shared: besides the IOC's IP2..IP6 it carries IP7, which the compare timer raises from the hptimer thread, and bit 63, the soft-reset request a device sets through CpuDevice::signal. A reset through the IOC (Machine::reset, a snapshot load or a rollback — all of them run power_on_devices) therefore also dropped a timer tick or a pending soft reset that landed while it ran, neither of which the IOC owns. The window is the gap between cpu.power_on(), which is entitled to clear the word, and ioc.power_on(), which is not.

The change

• power_on clears exactly the five lines update_interrupts drives (fetch_and, not store), so a reset no longer touches IP7 or bit 63.
• IOC_OWNED_LINES names that set once, with the ownership rule in its doc comment, so the reset path and the update path cannot drift apart.

Tests

• Unit test: power_on_clears_only_the_cpu_lines_the_ioc_drives: with every bit of the word set, a reset leaves everything except IP2..IP6, and the test pins IOC_OWNED_LINES to exactly those five, so IP7 (the compare timer's) and bit 63 (a soft-reset request) can never be folded into it.
• The test fails on main: assertion left == right failed: a reset clears the IOC's own lines and nothing else; left: 0, right: 0xffff_ffff_ffff_e7ff — the whole word went to zero.
• cargo test --lib: 1092 passed, 0 failed, 19 ignored (1091 before the change, i.e. the new test is the only difference). One unused_mut warning from net/testnet.rs:1292 is pre-existing on main.
• Booted NetBSD 10.0/sgimips on an Indy (R4400), headless, over the PROM's own TFTP path: the stock netbsd-INSTALL32_IP2x installation kernel served with --tftp-dir, then boot -f bootp()/netbsd from the Command Monitor, with an empty 512 MB SCSI disk and a CD-ROM attached. The kernel reports the R4400, its L1/L2 caches, ioc0 ... machine Indy (Guinness), wdsc0 ... WD33C93A, sq0 ... SGI Seeq 80c03, the emulated SGI, IRIS EMUL DISK and Sony, CDU-76S, and then starts sysinst (NetBSD-10.0 Install System). Log: netbsd-boot-fixed.log.
  • Same image on main: built without this change and booted the same way — same kernel banner, same device list, same installer (main-boot.log). This change is a no-op on the boot path.
  • A monitor reset — the Machine::reset path that runs power_on_devices, i.e. this function — leaves the CPU at 0xffffffffbfc00000, and the second TFTP boot reaches the installer again (netbsd-two-boots-and-reset.log).
• Built and tested with the default feature set on Linux only; the release-feature (lightning,jitv2,gpujit,idle-pause,hostgl) leg and the macOS/Windows legs were not run here.

Fixes #264

[techomancer]

thanks!