issues: the kernel-is-small track's input steps and exits deliver to the owner of input over the input port, and step 10's exit checks the kernel's streams deleted

[Japabu]

issues/the-kernel-is-small-interrupts-post-and-threads-wait.md no longer routes keyboard and mouse input through the kernel. Every clause of the track that presupposed a kernel feed is rewritten to the input ruling: userland owns the merge, drivers deliver to the owner of input over the input port, the kernel's merge is deleted, and the kernel's keyboard and mouse streams go when their last kernel source leaves, which step 10's exit checks.

The ruling is #899's, applied to usbd on #901 and recorded as #901's review NOTE 6. The fence was widened to the exits by the ruling on #901, recorded here. Nothing else in the file changes.

What changed, per clause:

• Step 10's body. It sent USB HID "to the keyboard claim", and a USB keyboard's transitions reached the kernel as usbd's keyboard claim. Now usbd delivers HID to the owner of input over the input port, as the ruling names it: Design: one userland input owner for the ARM desktop #868's inputserver once it lands, the compositor until then. The kernel's streams go when their last kernel source leaves. Step 10 depends on stage 7 step 2 landing first, because at step 10 usbd's is the kernel's last input source (the owner's ruling on this PR).
• Step 10's exit, third clause. It read "the keyboard's by stage 7 step 2's kernel test on usbd's claim", which has no subject once there is no keyboard claim (S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's design, "S5's exit"). The kernel's part now names the controller's claims alone. A new clause holds the owner of input to releasing a key usbd held at its death. Its test is toyos-input's: a connection's end releases every device it carried, and a gone releases only its own source. It is red with Merge's release on a connection's end removed, and red with gone releasing its whole connection: S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's two controls, exactly. This is the test S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's design names as "the owner's test that a key held at usbd's death is released".
• Step 10's exit, the kernel's input streams. The step's body says the kernel's streams go with their last kernel source, and stage 5's and 6.4's exits are step 10's, so the exit now checks it: keyboard.rs, mouse.rs and input_merge_test.rs gone from kernel/src, and no Keyboard or Mouse in sdk/abi's DeviceType or FileType. A grep fails it.
• Stage 7 step 2's body. It had ps2server feed "the kernel's keyboard and mouse streams so the merge with USB HID stays where it is". Now ps2server delivers to the owner of input over the input port, where userland does the merge. In ps2server: the i8042 driver leaves the kernel, and the owner of input merges every source #899's words, the kernel's merge (HELD, merged_buttons, SOURCE_EXISTS) is deleted, and its streams carry its own sources' records unmerged until usbd takes the last of them.
• Stage 7 step 2's keyboard-claim sentence. It read "A keyboard claim is refused while no source exists, which the supervisor's order of endowment then decides". It now says that ps2server's one kernel claim is the isa row, that it takes no keyboard claim, and that the owner reads its transitions from its input acceptor's connections.
• Stage 7 step 2's kernel test. This was a keyboard claim handed to the next claimant. It is now ps2server: the i8042 driver leaves the kernel, and the owner of input merges every source #899's mapping: the isa row whose holder dies is handed to the next claimant with none of the dead holder's records pending (isa_lines_reach_their_holder). It is red on a kernel whose release keeps the dead holder's record.

What I am unsure of: Step 10's exit cites the toyos-input test by what it checks, not by a name. #901's design does not name it yet.

The AMD bench's failure line is not filed again. issues/a-metal-job-that-fails-on-the-amd-bench-is-reported-as-failing-on-the-t14.md already records it at its site: Boot::job_passed, tests/common/metal.rs:490-498. Its exit is one a test can fail. #873's AMD judge line is one more reading of the same defect. This branch does not edit that issue, because an existing issue is edited only by the branch that takes it up.

Gates

• cargo run -- --ci host at ee3d8d62c, on origin/main at 6080e839d: EXIT=0, recorded as the log's last line (log).
• No production code, test or manifest changes, so no guest test is reached and no design round is owed.

Growth: one issue file, +29 −13 lines.

🤖 Generated with Claude Code

https://claude.ai/code/session_015xhaydXCVRNscoccHYFUWy

[Japabu]

Widening the fence: this branch also takes step 10 exit's third clause and stage 7.2's keyboard-claim sentence and kernel test, rewriting them to the input ruling (userland owns the merge; the kernel streams go with their last kernel source). Ruled on #901.

[Japabu]

Review, round 1: head 8728cbf97

Read against origin/main at 6080e839d. git log origin/main..8728cbf97 has three commits: 4068a0cda, the merge 1eaf1fbb0 and 8728cbf97. git diff --shortstat gives 1 file, +21 −13. All of it is in issues/the-kernel-is-small-interrupts-post-and-threads-wait.md, with 0 lines of production code and 0 of tests. The cargo run -- --ci host log at 8728cbf97 ends Host: 76 step(s), all green. Its FAILED lines all fall inside [ci] control arms, each of which reached its verdict. No guest test is reached, and no design round is owed for a records change. Each rewritten clause was checked against #899's body at 74aea28dc ("The track's step 7.2, and its exit", "The merge moves to userland whole", "Negative controls") and #901's body (section 3, section 8's toyos-input row, "S5's exit"). It was also checked against the rulings issuecomment-6137234392, -6137234651 and -6138651083.

The 7.2 kernel test matches #899's mapping. It names the existing T14 row isa_lines_reach_their_holder (tests/toyos.rs:968), whose dies role is the part this step owes, and the stated control matches #899's IRQ[row].clear() mutation.

BLOCKER

1. issues/the-kernel-is-small-interrupts-post-and-threads-wait.md:184-186 — the rewrite adds a deliverable to step 10: the kernel's keyboard and mouse streams go when their last kernel source leaves. No exit reads it. Step 10's exit (lines 188-211), stage 5's exit (which is "step 10's") and 6.4's exit (also "step 10's") can all be met with kernel/src/keyboard.rs, mouse.rs, input_merge_test.rs and DeviceType::Keyboard/Mouse (sdk/abi/src/syscall.rs:551,555) still in the tree. S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's S5 closes against this exit alone ("S5's exit is the track's step-10 exit as it stands once issues: the kernel-is-small track's input steps and exits deliver to the owner of input over the input port, and step 10's exit checks the kernel's streams deleted #904 lands"), so the deletion the ruling makes possible could land unchecked. This is the deletion S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's round-1 BLOCKER 2 required. Fix: add to step 10's exit something a build or grep can fail, for example keyboard.rs, mouse.rs and input_merge_test.rs gone from kernel/src, and no Keyboard or Mouse device class in sdk/abi.
2. issues/the-kernel-is-small-interrupts-post-and-threads-wait.md:182-183 and :313-314 — "delivered to the screen's owner" and "delivering to the screen's owner" state the ruling more narrowly than the owner gave it. The ruling names "the userland owner of input, which is the inputserver Design: one userland input owner for the ARM desktop #868 proposes, or the compositor until Design: one userland input owner for the ARM desktop #868 lands" (issuecomment-6137234392; -6137234651 for usbd). ps2server: the i8042 driver leaves the kernel, and the owner of input merges every source #899's own text says that Design: one userland input owner for the ARM desktop #868's inputserver "takes the acceptor". Once Design: one userland input owner for the ARM desktop #868 lands, the track as written sends ps2server and usbd to the wrong program. Fix: write "the owner of input", as step 10's exit already does at line 192.

NOTE

1. issues/the-kernel-is-small-interrupts-post-and-threads-wait.md:193-195 — "red with either release widened or removed" claims four mutations. The test the clause names checks two things: a connection's end releases every device it carried, and a gone releases only its own source. That test is red under three of the four mutations. It is not red when a connection's end is widened to release another connection's devices, because nothing in the clause's claim covers that case. S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's section 8 states two controls: Merge's release on a connection's end removed, and gone releasing its whole connection. Fix: either add "and nothing another connection carried" to the claim, or name S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's two controls exactly.
2. issues/the-kernel-is-small-interrupts-post-and-threads-wait.md:313-320 — ps2server: the i8042 driver leaves the kernel, and the owner of input merges every source #899's accepted text for this step says "the kernel's merge is deleted, and its streams carry its own sources' records unmerged until usbd takes the last of them". The rewrite drops the kernel merge's deletion (HELD, merged_buttons, SOURCE_EXISTS). "Where userland merges it" only implies it. Fix: state the deletion in the step.
3. PR body, Gates — the evidence log carries no exit code of its own. EXIT=0 appears only in the body, while the log ends at all green.

SEND BACK

[Japabu]

Ruling on the open point: the track's order puts 7.2 (ps2server) before step 10 (#900's order item 1), so the kernel's last input source at step 10 is usbd's, and the exit can be met. Say so in step 10's depends.

[Japabu]

Review, round 2: head a5c8e226e

Read against origin/main at 6080e839d. Since 8728cbf97 the branch adds one commit, a5c8e226e: +12 −7 in issues/the-kernel-is-small-interrupts-post-and-threads-wait.md. Over the whole branch it is 1 file, +26 −13, with 0 lines of production code and 0 of tests. The evidence log (ci-host-a5c8e226e.log) ends Host: 76 step(s), all green and then EXIT=0.

Round 1

• BLOCKER 1, no exit reads the deletion of the kernel's streams: CLOSED. Lines 191-194 add a check a grep can fail. Every name it cites exists at this head: kernel/src/keyboard.rs, mouse.rs and input_merge_test.rs; FileType::Keyboard/Mouse (sdk/abi/src/syscall.rs:551,555); and DeviceType's Keyboard/Mouse (device_classes!, syscall.rs:1319-1320).
• BLOCKER 2, "the screen's owner": CLOSED. Lines 183 and 317-318 now say "the owner of input", and the file no longer contains "screen's owner" anywhere.
• NOTE 1, the controls: CLOSED. Lines 197-199 now name exactly S1 of the xHCI move into usbd: a FAT entry that grows its file is written a cache flush after its chain and data, held by a module test that stops the disk inside a flush epoch #901's two controls.
• NOTE 2, the kernel merge's deletion: CLOSED. Lines 319-321 state ps2server: the i8042 driver leaves the kernel, and the owner of input merges every source #899's text, with HELD, merged_buttons and SOURCE_EXISTS.
• NOTE 3, the log's exit code: CLOSED. The log's last line is EXIT=0.

BLOCKER

1. issues/the-kernel-is-small-interrupts-post-and-threads-wait.md:182-186 — the owner's ruling (this PR's newest comment, posted after a5c8e226e was committed) is not applied. The ruling is that 7.2 comes before step 10, and step 10's depends must say so. At this head, step 10's body does not say it. Without it, step 10's new exit is only meetable by an ordering that nothing in the step states. That exit (lines 191-194) deletes keyboard.rs and mouse.rs and drops Keyboard/Mouse from the ABI. That can only be met once ps2server has taken the i8042, the kernel's other remaining source. Today the only place the ordering appears is the "What is left" list (lines 20 and 29-30), and nothing in step 10 itself states it. Change needed: add to step 10's body a sentence that it depends on stage 7 step 2 landing first, because at step 10 usbd's is the kernel's last input source. Cite the ruling on this PR.

NOTE

None.

SEND BACK