One assertion at two readings is one failure; an index is two

[Japabu]

The suite's ALONE: line re-runs one red alone and says whether the same
defect came back. It decided that by comparing the wide run's headline with
the isolated re-run's byte for byte, so an assertion that prints what it
measured wrote a different sentence every time it fired and the arm claimed
a second defect. Nightly ci run 35072262489 is the case: one assertion,
two boots, 0.303 s and 0.300 s, reported as red again on a DIFFERENT failure — it failed twice, on two assertions. That is the larger of the two
findings, so the defect sent an adjudicator looking for a second defect that
does not exist, and it did it to every timing, counting or sizing assertion
in the suite.

toyos_build::alone::same_failure decides it instead, on a rule derived
from what the two cases are rather than fitted to the sentences that
produced them: what differs between two readings of one assertion is a
measurement, and what differs between two failures is an identity. So only
what is unmistakably a measurement is taken out before the comparison — a
number carrying a unit of time or size (0.303 s, 1007 ms, 12 MiB), and
the time in a kernel record's stamp ([kernel 0.075 cpu0], which no two
boots write the same). Everything else stays, because it names which thing
was observed: slot 1 against slot 2, a port, an APIC id, a CPU, an
opcode, a register, a count with no unit. Merging two of those would lose
the larger finding, which is the defect this module exists to stop.

Every fixture in the module's tests is a line this tree prints, whole: the
floor and ceiling of tests/common/usb.rs's first-port window, the transport
break of kernel/src/drivers/xhci/wait/msc.rs with the controller's BDF
inside the slot (the line src/redlist.rs quotes off CI), the durability
sentence of tests/common/volumes.rs, the stall line of
kernel/src/heartbeat.rs. The units list holds ten spellings and each is one
an assertion here prints — µs, TiB and kB occur in no string literal in
tests/, kernel/src, userland/, src/, bootloader/, toyos/ or
toyos-abi/, and every entry only widens a merge that has to err toward
"different", so they are gone. every_unit_is_one_this_tree_prints holds the
list against the site that prints each.

What the scan does not reach

A scan closes the spellings it matches. A reading rendered with no unit
beside it reads as an identity, so one assertion printing two of them is
reported as two failures — the safe direction, and four assertions in this
tree are in it: check_physical's wake lateness rendered a second time as
pipeline depths, the dither floor's percentage, the tone peak's bare count,
and the log-drain verdict's bytes spelled out, which is not B. The module
header states the limit in one sentence, a test per shape asserts same=false
from that assertion's own wording, and
issues/build/the-alone-classifier-cannot-see-a-reading-with-no-unit.md
records the compromise with its owner, the four sites and two exit conditions.

The line itself gains one arm: a reproduction at two readings says red again, the same failure both times — the defect is real and prints the wide
run under wide: and the alone run under alone: rather than picking one.
One call decides it — byte equality is a case same_failure already answers,
so asking it first would be a second rule nothing tests.

The two checks

• Negative control, both arms, measured on this host at 6630db78
  (git status --porcelain empty before and after each, restored by a
  trap-guarded script):
  • The rule this replaces — same_failure is one == other:
    cargo build -p toyos-build --lib exit 0; cargo test -p toyos-build --lib alone:: exit 101, 8 passed 3 failed —
    one_assertion_at_two_measurements_is_one_failure,
    one_verdict_at_two_kernel_stamps_is_one_failure,
    an_index_is_an_identity.
  • The other extreme — reading returns the literal's length unconditionally,
    so every standalone number is a reading: cargo build -p toyos-build --lib
    exit 0; cargo test -p toyos-build --lib alone:: exit 101, 4 passed
    7 failed — an_index_is_an_identity (slot 1 against slot 2),
    a_skeleton_keeps_everything_that_is_not_a_reading,
    one_verdict_at_two_kernel_stamps_is_one_failure,
    a_reading_rendered_twice_keeps_the_copy_with_no_unit,
    two_percentages_are_two_failures, two_bare_counts_are_two_failures,
    bytes_spelled_out_is_not_a_unit.
  • The label swap on the new arm — the alone run's sentence printed under
    wide: and the wide run's under alone:, which is the mis-attribution the
    gate exists to stop said the other way:
    cargo build -p toyos-build --test toyos-build exit 0;
    cargo test --test toyos-build alone_line_reports_the_alone_run exit 1,
    FAIL alone_line_reports_the_alone_run: "the controller started at 0.303 s, …" is not the line's "wide: " run.
• Independent oracle: the recorded real failure. The fixture is the two
  sentences nightly ci run 35072262489 actually printed, which the
  classifier has to call one failure, and the pair from run 31424496450 —
  xhci_hid_break's endpoint count against its pointer delivery — which it
  has to keep apart. The usb-storage fixture is the line CI produced and
  src/redlist.rs records verbatim.

Green arms on this host at 6630db78: cargo test -p toyos-build --lib →
exit 0, 319 passed, 0 failed, 1 ignored; cargo test --test toyos-build alone_line_reports_the_alone_run → exit 0, PASS alone_line_reports_the_alone_run (52µs). Clippy, the root shape as
src/clippy.rs declares it with $ADOPTED expanded and -D warnings
(--workspace --all-targets --keep-going) → exit 0.

🤖 Generated with Claude Code

https://claude.ai/code/session_014iqcj4jDKpaiDX8B7CMvmK