Harness: metal boots taken whole by name, a stalled wait carries what the guest said, a guest test boots the console image

[Japabu]

Three harness clean-ups and one filed reading. Head e9185c18f, which holds origin/main 8a883a142 merged in (5939e8f19, clean) and one commit over it. The landing rests on CI at e9185c18f: host, toolchain / build and guest / suite have not run at this head when this is written; what has is under "Gates".

What changed, per decision

1. --metal takes named boots whole, in one invocation. No existing word did it: the one filter was a substring applied to members before chunking, so shared-2 was nobody's name.

• The suite takes every word that is nobody's value as a filter (it refused a second), and a run takes what any of them matches. Under --metal a word boot:<name> takes that boot whole: its shared members as the whole profile chunks them, and every row with an arm on it (which brings that row's other boots). cargo test --test toyos-build -- --metal --metal-readback <dir> boot:shared boot:shared-2 control_regs stages, and over the answered directory judges, those.
• metal::select (tests/common/metal.rs) is the one selection, and the chunking moves into it; shared_metal and c_corpus_metal lose their filter parameters. Four guards that selection made unreachable are deleted: the three "boot with no jobs" ones and run's "nothing to run".
• A word that takes nothing is refused by name in both suites (it would be dropped in silence beside one that takes something); boot: without --metal is refused at parse. parse refuses before any lock or build; under --metal a word that takes nothing is refused by select, which needs the corpus' built names, so that refusal comes after the shared binaries' build (63 s cold, metal-dead-boot.log). parse's header says so.
• A boot taken whole beside a member taken by name that lands on the same label is refused by batches as two shared boots are both named "…", which names the label and neither word. select's doc promises nothing about it.
• The line "the shared boots' members are not filtered by name" was false and is gone.

2. A stalled wait says what the guest said. await_guest's STALLED puts the last 40 lines the guest said during the wait under its sentence, with the count of what was cut. The sentence stays the first line, which is what the summary and the stall classifier read. window is the tail WaitVerdict::for_test already cut by, factored out and private to its module, so there is one cut and one place that holds its wording. No check holds it: that it keeps the tail under the count is read off eight moved lines.

• Deleted with it: The power-off says what it has to say above the console's last drain, and a one-CPU row reds without it #767's local capture on acpi_mediated_access's wait (.map_err(… "since its boot the guest said" …) and its comment, −3; the await_guest call ends in ? again). That capture began at the boot log's end, where await_guest's window begins, so a stalled wait there would have said its lines twice. acpi_lock_given_back_on_one_cpu runs the same function.
• Not changed: run_test_paced's ceiling (its capture is TestResult::serial), and the three callers that append their whole capture after the error (judge_virt_job, iommu_virtio_platform, and since A pipe watch can ask whether the other end is gone, and netstack decides a client has left by it #760 netstack_socket_churn's wait for the lease), whose captures open with the boot log the window does not hold. Read in the merged tests/toyos.rs: no other wait appends what the tail already says; the other await_guest caller, the_others_halt_first, is a bare ?.

3. console_image_boots: a headless guest boot of console/system.toml's image, on Profile::Metal (a panel for /system/bin/console to claim, the 16550 as console, -display none), waiting for console: ready and a clean console.

• What is not possible: a typed shutdown. The console's shell reads the i8042 and nothing else, and the tests that typed at it were deleted as red under host load: issues/the-console-loses-typed-keystrokes-under-host-load.md and issues/the-console-input-path-can-stop-after-a-ps2-overflow.md are open, their exits unmet.
• Why a guest test: no type or host test boots an image; a metal row cannot carry it, since metalimage::derive refuses a config that starts no runner and this image runs no job list and hands no machine back.
• Its predecessor screen_console_shell (deleted 958ada05e as a filed flake under other worktrees' suites) read the panel on flat 30 s ceilings; this reads the console only, under await_guest's liveness guard. Under the development machine's load it is unmeasured.

4. Filed: issues/the-power-offs-last-drain-gives-up-on-a-wire-whose-holder-cannot-run.md, the reading #767's review (comment 6058801505) named in serial::flush_final. Read from the code, not run. The reading turns on the seal: preemption goes off where xhci::seal_shut takes the controller lock, nothing keeps klogd from running between drain_inline and it, and the exit's actuator stages klogd inside its hold at the seal. Its citations were read, by me and by review round 2, at 6e3cb894f, whose kernel was d78350c71's; the merge brought #767's and #766's kernel changes under them and the line numbers were not read again at this head.

Review round 1, per finding

finding	answer
BLOCKER, step 6 of ceiling_self_check	deleted (−14), and mut-window.patch with it; window is private; rg 'last 40|lines in it' tests src finds tests/common/qemu.rs:410 alone
NOTE, "nothing to run" guard	deleted (−4)
NOTE, the same-label refusal names neither word	the smaller: the sentence in select's doc is dropped (−2)
NOTE, the issue's claim	rewritten on the seal, as above; the refused-seal arm is said not to be covered
NOTE, the issue's citations	console.rs:94, machine.rs:94, serial.rs:278-279, xhci/mod.rs:1550
NOTE, the body's gates; parse's header	below; the header now says which refusal comes after a build
BLOCKER, no host, no guest suite	closed at 6e3cb894f by CI run 37773729360 (round 2 judged it so); owed again at e9185c18f

Review round 2 (comment 6059791981, LAND AFTER NAMED CHANGES), per finding

finding	answer
NOTE, #767's local capture doubles the tail once both are on main	#767 landed; origin/main 8a883a142 merged (5939e8f19), the .map_err and its comment deleted in e9185c18f (tests/toyos.rs +1 −3). The two tests that share the function ran green at this head, below
NOTE, CI built the merge with ed80f9099; #766 and #760 since	merged here; cargo metadata --locked and toyos-checks green on the merged tree, below. host and guest / suite at e9185c18f are the measurement of the rest
NOTE, the body	this text: the head and origin/main; CI run 37773729360 under "Gates", the round-1 table and "Unsure"; three whole-capture callers

Friction removed against lines added

git diff --shortstat origin/main...e9185c18f: 7 files, +302 −109.

item	lines	of which checks	replaces
1 selector	+196 −94	+81 −13 (tests/checks.rs, tests/checks/metal.rs, testargs' tests)	staging 30 images to delete 27, or nine one-substring stagings and nine T14 boots for one pull request; a false line
2 stall tail	+23 −15	none	an instrument patched into a test per stalled wait; #767's local map_err, deleted here
3 console boot	+21	the test	three hand boots that opened QEMU's window
issue	+62

Harness code outside checks: item 1 +115 −81, item 2 +23 −15.

Gates

At e9185c18f, on the development machine, one script (176 s), each line the command's own exit; the tree was clean after it:

command	exit	deciding line
cargo metadata --locked	0
cargo test --test toyos-checks	0	test result: ok. 37 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out
cargo test --test toyos-build -- acpi_mediated_access acpi_lock_given_back_on_one_cpu	0	PASS acpi_mediated_access (3s), PASS acpi_lock_given_back_on_one_cpu (138s) (135 s of that the cold build of its binary), test result: ok. 2 passed, 2 total; both under TCG, each printing acpi: the Global Lock given back for a holder that left it taken (the machine is stopping)

These two are green arms only: what the deletion changes is the text of a red, and no stall was provoked on this wait. The tail itself was read red under mut-stall.patch on console_image_boots, below.

Not run at e9185c18f: cargo run -- --ci host, toolchain / build, the whole guest suite. They are CI's.

At 6e3cb894f, CI run 37773729360 (pull_request, the merge of 6e3cb894f into ed80f9099, all three jobs success):

job	id	deciding line
host (cargo run -- --ci host, clippy in it)	113299322077	[ci] Host: 78 step(s), all green
toolchain / build	113299322574	success
guest / suite, under KVM	113300721222	PASS console_image_boots (3s); test result: ok. 33 passed, 33 total

On the tree 6e3cb894f commits, on the development machine (run before that commit, over 795a74859; the script's diff of the tree before and after is identical):

command	exit	deciding line
cargo test --lib testargs	0	test result: ok. 17 passed; 0 failed; 0 ignored; 0 measured; 387 filtered out
cargo test --test toyos-checks metal_words under mut-selector.patch	101	test checks::metal_words_take_rows_members_and_whole_boots ... FAILED, left: Ok(("", "")), right: Ok(("", "shared=test_rs_a1+test_rs_a2")); 0 passed; 1 failed

At 23857ec9d. Between it and 6e3cb894f the harness moved by a doc comment in src/testargs.rs, pub fn window made fn window, a doc sentence and the deleted "nothing to run" guard in tests/common/metal.rs, and the deleted check step; since then by the merge and this round's −3. None of these rows was run again:

command	exit	deciding line
cargo test --test toyos-build -- console_image_boots machine_shutdown	0	test result: ok. 3 passed, 3 total: console_image_boots (3s), machine_shutdown, and machine_shutdown_short_stop, which the second word is part of
… -- console_image_boots no_such_test	1	No test matches filter "no_such_test"
… -- boot:shared	1	boot:shared names a boot of the metal profile, and a run without --metal has none … add --metal
… -- console_image_boots under mut-stall.patch	1	STALLED: waiting for the console to take its panel — it went quiet, (the last 40 of the 59 lines in it), the tail's last line console: ready 1920x1080 (240x67 cells) …; the log ends exit status: 1
… -- console_image_boots under mut-console-image.patch	1	the same sentence; the tail ends at logkeeper: with no console: ready; the log ends exit status: 1
… -- --metal --metal-readback <dir> boot:shared-99	1	boot:shared-99 names no boot of the metal profile, whose boots are {…}, printed 11:35:09 by a run whose first build line is 11:34:06
… -- --metal --metal-readback <dir> boot:shared-2 boot:ccorpus control_regs	2 (staged)	1 registration(s) and 159 shared member(s) over 3 boot(s): ccorpus: 137 job(s), shared-2: 22 job(s), testcases: 7 job(s); the log ends exit status: 2

No T14 run was made or is asked for; the staged images were removed. The three mutation patches are in the first comment; git apply --check of each exits 0 at e9185c18f, and none was run again here.

When #764 meets this

#764 (wt/toyos-amls2, open): issues/no-harness-path-boots-the-console-image.md is not on main, and this does not meet its exit as written (a shutdown typed and guest-shutdown read). What this gives it: the hand boots are replaced, and console_image_boots takes one await_marker on its hand-over line, after which deleting acpiserver from console/'s [boot] start reds a test. On main the console image starts no acpiserver, so that line cannot be asserted here. The issue's body and exit are the orchestrator's to narrow or keep.

Unsure

• console_image_boots under a loaded development machine, the condition its predecessor was deleted under: unmeasured. Its readings are one under TCG, three tests wide on a quiet machine, and one under KVM inside CI's whole suite (job 113300721222).
• The merged tree under the whole suite and --ci host: CI at e9185c18f. Run 37773729360 measured this branch over ed80f9099, before The AML interpreter is walked in declaration order and says what a call took #766, A pipe watch can ask whether the other end is gone, and netstack decides a client has left by it #760, The power-off says what it has to say above the console's last drain, and a one-CPU row reds without it #767 and munmap and a FIXED mmap take only the length the mapping was asked for #765.
• The issue's reading is of the code alone. That the stop's thread is preemptible up to the seal rests on drain_for_the_stop parking on the wire and on nothing read between there and the seal taking preemption away.
• Untried: once acpiserver is in the console image, a QMP power-button press (as acpi_power_button makes) may reach guest-shutdown with no keyboard, which would be a headless power-off of that image though not its shell's shutdown.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RvnWQFcMuGqTHYhvSnTe8A

[Japabu]

The three mutation patches the body's gates name. Each was made against 795a74859 and applies unchanged at 6e3cb894f (git apply --check exit 0); mut-window.patch left the record with the check step it turned red, deleted in review round 1. Each was checked with git apply --check, applied, run, and reversed in the same script; the tree was clean before and after.

mut-selector.patch

diff --git a/tests/common/metal.rs b/tests/common/metal.rs
index 1217c55b1..ef147fd3b 100644
--- a/tests/common/metal.rs
+++ b/tests/common/metal.rs
@@ -204,7 +204,7 @@ pub fn select(
 
     let named = |name: &str| all || names.iter().any(|word| name.contains(word));
     let mut taken: Vec<SharedBoot> =
-        whole.iter().filter(|boot| boots.contains(&boot.boot.as_str())).cloned().collect();
+        whole.iter().filter(|_| false).cloned().collect();
     let by_name: Vec<SharedBoot> = shared
         .iter()
         .cloned()

mut-stall.patch

diff --git a/tests/toyos.rs b/tests/toyos.rs
index cc34a7854..e8f1d5998 100644
--- a/tests/toyos.rs
+++ b/tests/toyos.rs
@@ -1774,7 +1774,7 @@ fn console_image_boots() -> Result<(), String> {
         BootOptions { profile: qemu::Profile::Metal, ready_marker: bootlog::COMPLETE, ..Default::default() },
     );
     let mut console = format!("{}\n", qemu.boot_log());
-    await_marker(&mut qemu, &mut console, "console: ready ", "the console to take its panel")?;
+    await_marker(&mut qemu, &mut console, "console: never ", "the console to take its panel")?;
     serial::Serial::named("the console image's boot", console).must_be_clean()
 }
 

mut-console-image.patch

diff --git a/console/system.toml b/console/system.toml
index 65be9a1a6..57ba9737b 100644
--- a/console/system.toml
+++ b/console/system.toml
@@ -16,7 +16,7 @@
 assets = ["assets"]
 
 [boot]
-start = ["logkeeper", "diskserver", "fileserver", "console"]
+start = ["logkeeper", "diskserver", "fileserver"]
 
 # `/system/bin/console` owns the screen and provides the surface its shell and the
 # shell's children inherit. No compositor here, so nothing receives one.

[Japabu]

Review of 795a74859 against origin/main 47cbd2e5b, round 1. Read only; nothing was run. git merge-tree --write-tree origin/main 795a74859 exits 0 (tree d4132dc65677); against #767's e3a2022d1 and #764's 11031d4c7 it also exits 0.

Net: 8 files, +305 −101 (git diff --shortstat origin/main...795a74859). Kernel and userland: 0. Harness outside its checks (src/testargs.rs's parser, tests/common/, tests/toyos.rs): +156 −88. Checks (tests/checks*, testargs's tests): +95 −13. Issue: +54.

BLOCKER

• Evidence — no cargo run -- --ci host at this head, and the guest suite is unread — the body has cargo test --test toyos-checks (0, 37 passed) at 795a74859 and cargo test --lib testargs at 23857ec9d; clippy at CI's version, the source and issue gates and every other host suite are unmeasured, and console_image_boots has been read once, under TCG, three tests wide on a quiet machine. Closes with no change to the branch when host, toolchain / build and guest / suite are green at the landing head; guest / suite is the first reading of the new test under KVM and inside a whole suite's load, which is the condition its predecessor screen_console_shell was deleted under (958ada05e).
• tests/checks/qemu.rs:291-304 — step 6 of ceiling_self_check is a test of what a reader checks, and asserts on the tail's text — window (tests/common/qemu.rs:406-415) is eight lines moved verbatim out of for_test; that it keeps lines[lines.len() - kept..] and heads it with the count is read off the diff, and mut-window.patch is a mutation a reader of the diff catches. Delete the step (−14); nothing else in the tree then holds the tail's wording, which is what the stall tail was asked to be.

NOTE

• tests/common/metal.rs:972-975 — the "nothing to run" guard is dead and the diff edited it instead of deleting it — select refuses every word that takes nothing and no word takes the profile, so main cannot reach it, and tests/checks.rs:943 passes one row; it is the fourth of the three guards the body deletes on this same argument (−4).
• tests/common/metal.rs:205-216 — a boot taken whole beside a member taken by name that lands on the same label is refused by batches as two shared boots are both named "shared", which names neither word — say in select, where the words are still in hand, which boot: word and which name word met, or drop the sentence in select's doc that promises the refusal. No test is asked for.
• issues/the-power-offs-last-drain-gives-up-on-a-wire-whose-holder-cannot-run.md:35-38,58-59 — "a klogd descheduled inside its hold when the stop's thread passes drain_inline is never run again" is stronger than the cited lines: preemption goes off at seal_shut (machine.rs:132), and nothing cited keeps klogd from running between :112 and :132. The point the reading turns on, and the one the exit's actuator must stage, is the seal, not drain_inline.
• The same file, citations — console.rs:90 is drain_inline's signature and the quoted let is :94; machine.rs:95 is :94; the SleepLock's sentence is serial.rs:278-279; xhci/mod.rs:1561 is seal_shut and the quoted sentence is :1550. Prose.
• PR body, Gates — commands, exits and a deciding line each, not the logs; the logs in the bench's scratchpad say what the table says (mut-stall.log and mut-console-image.log end exit status: 1, metal-stage.log exit status: 2). The guest rows are from 23857ec9d: git diff --stat 23857ec9d 795a74859 is tests/checks.rs and the issue file, neither an input of toyos-build, so I take them as this head's. Prose.

1. The selector, ruled

• boot:<name> stands, as a prefix and not a flag or a merged namespace. A boot's label is a third namespace beside row names and member names: shared is a boot and a substring of member names, so one vocabulary cannot carry both without the word saying which it is. A --boot flag in SUITE would be read once (Value::Next refuses a second use), and Each would be a new value kind in the harness's table for one reader. The substring filter is not replaced by it: the implementer's --metal-readback <dir> <row> needs row names. Several filter words is the general half and serves the QEMU suite too.
• --metal without --metal-readback. Every exit before metal::run read: parse refuses an unknown - word, an empty word or boot: alone, a word that is a flag's name, and boot: without --metal (src/testargs.rs:49-71,96-108), and exits before any lock; select refuses a name word that is part of nothing and a boot word that names nothing, each beside words that do take something (tests/common/metal.rs:193-203), and main exits 1 on either (tests/toyos.rs:4928-4934); batches' refusal returns Red above the first build and the only drive (metal.rs:958-964,1053). all is true only with no positional at all (metal.rs:191), so no refused or mistyped word reaches "everything": boot:shard is refused by name with the boots there are, bootshared and Boot:shared are dead name words, --metal-readback boot:shared is an offline run into a directory of that name. The check holds both refusals and (&[], &["shared"]) taking one boot, so all losing its boots.is_empty() reds it. What stays as before: --metal shared is a substring over names and takes whatever contains it.
• A dead word under --metal costs the shared binaries' build first (metal-dead-boot.log: 63 s cold), because select needs the corpus' built names; parse's header still says a refusal costs a message. Not asked to change.
• Chunking. A boot taken by word is an element of sized(shared) over the unfiltered profile (metal.rs:198,206-207), which is the value a run with no word passes on (by_name is then shared itself), and it brings every row with an arm on it, so its batch's job list is the whole profile's in the same order: a judge of that readback reads the members the full run would. Not so for what comes in tow: a row taken that way brings its other boots carrying that row's jobs alone, and a member taken by name is still chunked as it comes, under shared or ccorpus whatever chunk the profile has it in. Both are staged under the profile's labels with fewer members, as a name word always did; request.txt and the judge print each boot's member count, and that count is what tells them apart.
• Growth. Harness +113 −76 and checks +81 −13 for it: accepted, less the guard above. words_take_rows_members_and_whole_boots stays: what a --metal line flashes is decided by the composition of the filter, sized over two inputs and the row-arm join, and its refusals are the ones that keep a typo off the machine.

2. The stall tail, ruled

Bounded at 40 lines with the count of what was cut; one cut (window) and one site (await_guest); the sentence stays the first line, and the summary's line and stalled() read it as before (mut-stall.log: the failures: line is the sentence alone). window has two callers, WaitVerdict::for_test and await_guest, and both take a QemuInstance's capture: no metal readback passes through it, so nothing the T14 says about itself can ride it; a QEMU guest's lines name QEMU's machine, the commit and the clock. iommu.rs:77 and judge_virt_job now print the tail and then their whole capture; their captures hold the boot log the window does not, so they stay. Whichever of this and #767 lands second deletes #767's map_err on acpi_mediated_access's wait in its own diff: after both are on main it is a sibling.

3. console_image_boots, ruled

• Tier. The body's reason holds by the code: metalimage::derive refuses a config that starts no runner (src/metalimage.rs:73-75, "a T14 boot that does not end itself parks on a console nothing is on"), so no metal row can carry console/system.toml as the harness stands; no type or host test boots an image.
• Registration. MACHINE_TESTS with its reason, and run_machine_test's arm; registered holds the name against every other. No new boot config: console/system.toml is already in ALL_CONFIGS (src/build.rs:3097), and no feature, parameter or kernel build is added (the capture's census: 1 kernel build(s): [""]).
• Cost per run. One ROOT of the console image (2 s once its crates are built) and one boot, PASS (3s) under TCG; a red costs 18 s.
• Can it fail on what it claims. mut-console-image.patch (the image starts no console): STALL, exit 1, the tail ending at logkeeper: with no console: ready. It asserts content and completion and no duration.
• The owner's ruling on test programs that run only on the T14 is not moved by this diff: no METAL row, test binary or config changes tier, and CI runs no --metal. What this boots is a shipped image mode, not a test program, and it ran under QEMU until 958ada05e; whether that ruling reaches an image built for the T14's panel is the owner's to say, and I do not extend it.

4. The filed issue, ruled

Frontmatter and slug meet issues/README.md; owner named and existing; the exit is a guest test that reds on today's drain. flush_final's bounded try_wire and the black-box line, klogd's preemptible hold, drain_inline declining a held wire and seal_shut forgetting its guard are as the code has them at this head (the kernel is unchanged since d78350c71). The two NOTEs above are what is not. It is not in #767's diff, so nothing files it twice.

5. The proposed rule

Declined for implementer.md. Its sentence there (--metal --metal-readback <dir> <row>) is still true, and "one invocation per pull request" is something several words make possible, not something a reader can violate unseen. The usage's home is where it already is: testargs::BOOT's doc, --metal --list printing the labels, and the dead-boot refusal listing them; a review or brief that wants a boot whole writes the command it wants run.

The landing

host, toolchain / build and guest / suite, none of which has run (the pull request is a draft and all three report skipped). #767 and #764 merge clean with this textually; the map_err above is the one thing the second of this and #767 owes.

SEND BACK

[Japabu]

Review of 6e3cb894f against origin/main b26542c26, round 2. Read only; nothing was run. Reviewed: git diff 795a74859 6e3cb894f (one commit, 5 files, +41 −53), the issue file whole against the kernel at this head, CI run 37773729360 and its saved logs, the body.

Net: 7 files, +301 −106 (git diff --shortstat origin/main...6e3cb894f). Kernel and userland: 0. Checks (tests/checks*, testargs' tests): +81 −13. Harness outside its checks: +158 −93. Issue: +62.

Round 1's BLOCKERs

• Evidence: host, toolchain / build, guest / suite unmeasured: CLOSED. Run 37773729360, event pull_request, headSha 6e3cb894f0ef…, all three jobs success (113299322077, 113299322574, 113300721222). The checkout in both saved logs is Merge 6e3cb894f… into ed80f9099…. host: cargo run -- --ci host ends [ci] Host: 78 step(s), all green, with toyos-checks at 37 passed; 0 failed inside it; the log's 114 FAILED lines all stand under a === [ci] control … heading and each control ends N verdict(s) reached, so they are the negative controls reaching red, not a red step. guest / suite: /dev/kvm opens, -accel kvm; BUILT x86_64 ROOT of console, for console_image_boots (2s), PASS console_image_boots (3s), test result: ok. 33 passed, 33 total (375.0s; workers: 249s building, 126s testing). That is the new test's first reading under KVM inside a whole suite, and it is one reading: it says nothing about the loaded development machine its predecessor was deleted on, which the body already says is unmeasured.
• Step 6 of ceiling_self_check: CLOSED. tests/checks/qemu.rs −14, window is fn again (tests/common/qemu.rs:406); in the tree merged with origin/main, git grep 'last 40\|lines in it' over tests and src finds tests/common/qemu.rs:410 alone, and window( has its two callers in that file (:463, :595). host green at this head is the build without the step.

Round 1's NOTEs

• "nothing to run" guard: closed, tests/common/metal.rs −4; select returns Err on every word that takes nothing and all is the only way to an unfiltered profile, so run no longer holds an arm nothing reaches.
• Same-label refusal: closed by the smaller option; select's doc promises nothing, batches still refuses.
• The issue's claim: closed. As rewritten it is true by the code at this head: preemption leaves the stop's thread where seal_shut forgets the controller guard (xhci/mod.rs:1562-1566), drain_for_the_stop parks on the wire (console.rs:113-114), nothing between machine.rs:112 and :132 takes preemption, the None arm forgets nothing, and the exit's actuator is now placed at the seal. Every citation read: serial.rs:131, :263, :264, :278-279; console.rs:94, :417; machine.rs:69, :94, :96-:109, :112, :132; xhci/mod.rs:1550, :1565; power.rs:21, :53: each is the line it quotes. Frontmatter, owner and exit meet issues/README.md. Neither origin/main nor The power-off says what it has to say above the console's last drain, and a one-CPU row reds without it #767's ad450251c carries this slug or this reading.
• parse's header: closed (src/testargs.rs:44-49).
• The gates' logs: host-checks.log, host-testargs.log and mut-selector.log say what the body's table says (37 passed; 17 passed, 387 filtered out; metal_words_take_rows_members_and_whole_boots ... FAILED, left: Ok(("", ""))), and tree-before.diff and tree-after.diff are the same bytes.

BLOCKER

None at this head.

NOTE

• tests/toyos.rs (The power-off says what it has to say above the console's last drain, and a one-CPU row reds without it #767's acpi_mediated_access, .map_err(|why| format!("{why}\nsince its boot the guest said:\n{}", …)) and its comment line) — owed by this branch, because The power-off says what it has to say above the console's last drain, and a one-CPU row reds without it #767 lands first — git merge-tree --write-tree of 6e3cb894f exits 0 against origin/main (3dfb319d7d76), against origin/wt/toyos-poweroffline ad450251c (cc01b6b60004), and against the two merged together (bacc21489fb3), so the queue would land it as it stands and main would then say a stalled wait's lines twice on that wait: await_guest's tail and, under it, the same capture from the boot log's end. That is the sibling round 1 named. Not to be armed as it stands. Once The power-off says what it has to say above the console's last drain, and a one-CPU row reds without it #767 is on main: merge origin/main, delete the .map_err and its comment (−3, the ? back on the await_guest call), and let host and guest / suite run at that head; acpi_lock_given_back_on_one_cpu shares the function and needs nothing else. Armed before that deletion, the merged tree is a BLOCKER.
• CI at 6e3cb894f built the merge with ed80f9099; origin/main has since taken The AML interpreter is walked in declaration order and says what a call took #766 and A pipe watch can ask whether the other end is gone, and netstack decides a client has left by it #760. Read in the merged tree: A pipe watch can ask whether the other end is gone, and netstack decides a client has left by it #760's netstack_socket_churn adds a MACHINE_TESTS row, a run_machine_test arm and one await_marker(…).map_err(|e| format!("{e}\n{console}")); none meets a changed signature (shared_metal(), c_corpus_metal(&c_bins) and metal::select(filters, &parsed.boots, …) have this branch's callers only), and that map_err appends a capture that opens with the boot log, which the window does not hold, so it stands as judge_virt_job's and iommu_virtio_platform's do. The run owed above is the measurement of this.
• PR body, prose, the orchestrator's: origin/main is b26542c26, not ed80f9099; "Gates" and the round-1 table still say host, toolchain / build and the guest suite have not run and the BLOCKER is open, and "Unsure" still says KVM and CI's clippy are unread: run 37773729360 answers all of it and belongs there with its job ids and deciding lines; "the two callers that append their whole capture" are three on main since A pipe watch can ask whether the other end is gone, and netstack decides a client has left by it #760.

The owner's ruling on fewer, fuller T14 boots

Nothing in this diff has to be undone for either step. Step A (rows moved onto existing boots): select already takes, for a boot: word, every row with an arm on that label beside the shared chunk of the same name, so a row moved onto shared is taken with it by the word that exists. A chunk's label is positional (chunk_name: shared, shared-2, …), so moving members changes which members boot:shared-2 names; the refusal lists the labels there are and --metal --list prints them, which is what makes that safe. Step B (the in-guest runner bounds each member and carries on): no file of the runner is in this diff; where it changes how many members a boot holds, sized is now the one place the profile is chunked, called from select alone, where on main the chunking sat behind two filter parameters.

LAND AFTER NAMED CHANGES