issues: the kernel's drivers and threads, and the host binaries the build rests on, each with its owner

[Japabu]

The owner asked which standing debts remain besides the crate sprawl (#604) and the C corpus. This PR moves no code: it updates the issues that own each debt, files one issue for a debt nothing owned, and corrects one tree fact in CLAUDE.md. Issues only; no inventory or strategy document is recorded.

Changes

• issues/kernel/the-kernel-still-creates-threads.md: K2 is deleted because Kernel: a kill never waits on its victim — the last thread out tears its process down #549 did it. K6 waits on K4 and K5. K4 carries the constraint that one console wire's driver stays in the kernel (log::console::drain_inline, serial::panic_flush), and no longer cites a driver-model design.
• issues/kernel/every-driver-is-still-in-the-kernel.md: item 1 is re-scoped to the HDA and virtio-sound stubs and virtio-gpu, the ABI that retires with them, and why GOP stays.
• issues/kernel/the-kernel-is-small-interrupts-post-and-threads-wait.md: step 10 notes the MSI-X-in-BAR-0 refusal and that where the T14's xHCI and NVMe keep their tables is unmeasured.
• issues/build/python-and-cc-are-declared.md: what removing Python, cc, and CMake with Ninja takes, without point-in-time counts or versions.
• New: issues/build/the-owners-flash-script-runs-diskutil.md; its exit is rg -l "diskutil|plutil" over the tree outside rust/ finding nothing.
• CLAUDE.md (owner-approved tree fact): "four macOS FAT tools" becomes "two macOS FAT tools"; the tree has newfs_msdos and hdiutil only. src/sourcegate.rs loses the clause "Two, where CLAUDE.md says four", which that edit makes false. Nothing else in CLAUDE.md changes.

Gates

• cargo run -- --ci host: EXIT=0 on the head.
• No guest test applies.

Unsure

• Where the T14's NVMe and xHCI keep their MSI-X tables is unmeasured. The step for the orchestrator, read-only: lspci -nn for the two BDFs, then sudo lspci -vvv -s <bdf> as root for each. Without root the MSI-X capability's Table: BAR=N offset=... line is hidden as <access denied>. The metal loop's sudoers rule does not list lspci, so this needs an interactive root shell.
• every-driver-is-still-in-the-kernel.md still says USB HID cannot move without moving "the boot block device". That predates ROOT in memory; only the owner can rule on it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01U6SVYFkdvV2t38KzNrESxs

[Japabu]

Head 4223c5f. CI host on the head: SUCCESS (a second, earlier host run SKIPPED). Diff: 70 insertions, 10 deletions, issue files only; nothing to guess-build. No guest claim, so no metal reading owed.

Spot-checks against the tree, all true: MAX_KERNEL_TASKS = 2 (kernel/src/sched/kthread.rs:22); the only kthread::spawn sites are log/console.rs:66 and iod.rs:15; the pass calls xhci::poll_if_pending, keyboard_controller::service, pcidev::drain_pending, AUDIO_WATCH.post and watchdog::feed (sched/driver.rs:512,665,666,679,684); drain_inline (log/console.rs:91) and serial::panic_flush (serial.rs:139) exist; newfs_msdos and hdiutil are the only two FAT tools (toyos-fat32/tests/common/mod.rs:527,604, sourcegate.rs:601,608); flash.sh runs diskutil, plutil, shasum, dd and sudo, and the README runs diskutil; download_beta_toolchain exists in rust/src/bootstrap; nvme.rs is 908 lines and hpet.rs 97; the referenced issue files exist. The Command::new("./x") spawn is at toolchain.rs:631,633.

BLOCKER

• none.

NOTE

• PR body, section A step 1 — the T14 step is not concrete or safe to run as written. lspci -vv as a non-root user prints Capabilities: <access denied>, so the MSI-X Table: BAR=… offset=… line is absent. The metal loop's sudo is limited to the rendered /etc/sudoers.d/toyos-metal rule (src/metal.rs:7-9), which does not list lspci. State the command (sudo lspci -vvv -s <bdf> for the NVMe and the xHCI, with the BDFs from lspci -nn), that it needs an interactive root shell, and the field to read (MSI-X: … Table: BAR=N). It is read-only, so safe once root exists.
• PR body, "Unsure" — the CLAUDE.md rewording is only "the handoff proposes one sentence"; it is not in the PR or in the body. Nothing to judge as written. My recommendation is below.
• PR body — 8.1 s EXIT=0 (bootstrap build) and the two rust-lld results give no command and no log. The brief requires command, exit code and log for a measured claim. The claims are used as evidence in the issue text, so add the commands or drop the numbers.
• PR body — the 21-row inventory, the 7-step order and the stage list are inventory and strategy that will be the merge commit's record and live in no issue. The order is a proposal that the owner did not ask to be recorded. Consider cutting it to the measured table and the issue edits.
• issues/kernel/the-kernel-still-creates-threads.md K4 — the PR body says the "driver-model design" is not written anywhere in the tree, yet K4 still says "the owner-approved driver-model design moves the console to logd". The branch had the evidence and left a claim it just called unsupported.
• issues/build/the-owners-flash-script-runs-diskutil.md — Exit says "or by nothing", which cannot be checked. Name a check (for example, rg diskutil over the tree outside rust/ finds nothing).

REMOVE

• issues/build/python-and-cc-are-declared.md — llvm/CMakeLists.txt:1016, the 726/78/740 counts, LLVM 22.1.0, MacOSX14.4.sdk, MacOSX27.0.sdk, 1.98.1 and 9c3eea44: point-in-time numbers and cites that rot with every fork bump or SDK update. Keep the fact, delete the numbers.
• issues/build/python-and-cc-are-declared.md — "Nobody has run it end to end, and whether to try is the owner's call": narration; the ruling is already cited in the same file.

CLAUDE.md rewording recommendation. Yes, change it, and only the number: "four macOS FAT tools" to "two macOS FAT tools". The tree has two (newfs_msdos, hdiutil), and sourcegate.rs:611 already records the mismatch in prose that will say something false after the fix, so that sentence goes in the same edit. It is shorter by nothing and true. A longer rewrite is not needed: the ledger in src/sourcegate.rs already lists the tools. Not needed is the flash-script addition; that belongs in its issue, which the PR files.

LAND AFTER NAMED CHANGES

🤖 Generated with Claude Code

https://claude.ai/code/session_01U6SVYFkdvV2t38KzNrESxs