x86: removing non-error log records from the dispatch path exposes a doubled bsh prompt after init start (production-profile gate 5/9 red)

[ryanbreen]

Removing the whole non-error log:: record class from
kernel/src/interrupts/context_switch.rs reddens the x86 production-profile
gate on 4 of 9 boots with a console-prompt signature that 8 baseline boots and
13 boots of two narrower arms never produced. The removal is therefore not
shipped by #775, and this issue records what it costs and what the evidence is.

Signature

docker/qemu/run-x86-prod-profile-boot-test.sh aborts on one of its three
prompt checks:

:942   test "$PROMPT_AFTER" -gt "$PROMPT_BEFORE"
:952   test "$PROMPT_BEFORE" -eq 1
:953   test "$PROMPT_AFTER" -eq 2

The failing boots report a console prompt count of 3 or 0 where a healthy boot
reports 1 -> 2. The 3-prompt face is a DOUBLED prompt emitted immediately
after init announces itself. Verbatim, from the committed specimen
docs/planning/green-program/sockets/serials/775/round3/case-d-broad-removal/specimens/prod-r3head-b/boot3/serial_user.txt
lines 75-78:

[init] Breenix init starting (PID 1)
[SW]<K>[SW]<T><B>[SW]<K>
breenix> 
breenix> [SW]<T><B>[SW]<K>...

The 0-prompt face is specimens/prod-r3head-b/boot4/: no prompt in the whole
capture, so PROMPT_AFTER never exceeds PROMPT_BEFORE.

A/B, one host, one QEMU at a time

arm	boots	pass	fail	failure sites
51d7468f, the pre-#775-round-3 head	8	6	2	:1016 x2 (bsshd started != 1)
the census on COM2 plus the idle hook, 33 of 33 records kept	5	5	0	--
+ only the 3 records finding F15 named removed (what #775 ships)	5	5	0	--
+ the 22 non-error records removed	9	4	5	:953 x2, :942 x1, :952 x1, :1016 x1

The prompt signature is 4 of 9 on the broad arm and 0 of 13 on the two narrower
arms. :1016 is a separate, pre-existing flake that appears on the baseline
too.

Recipe: build the zero-feature production profile
(cargo build --release --bin qemu-uefi), run
docker/qemu/run-x86-prod-profile-boot-test.sh repeatedly, and read the
failing command: and console prompt: lines out of each transcript.

Hypothesis

The removed records are log::trace!/log::debug!/log::info! calls inside
the dispatch functions -- switch_to_thread, restore_userspace_thread_context,
setup_first_userspace_entry and their neighbours. Deleting them changes the
timing of the dispatch path, and that timing change exposes a race between init
starting and the console writing its prompt. The logging was load-bearing for
something. What that something is, is not known: the #775 rounds measured the
regression and declined the removal rather than chasing the mechanism.

Note that CLAUDE.md forbids serial output in exactly these functions, so the
19 surviving non-error records there are a standing violation that cannot be
removed until this is understood. That is the reason this issue matters beyond
the one gate.

Committed evidence

• docs/planning/green-program/sockets/serials/775/round3/case-d-broad-removal/README.md
  -- 22 summary rows, one per uncommitted boot, each with its PASS count, abort
  site and prompt line, plus the commands that produced them.
• .../case-d-broad-removal/specimens/prod-r3head-b/gate-{3,4,5}.txt with
  boot{3,4,5}/serial_{kernel,user}.txt -- three of the four prompt-signature
  boots in full.
• .../case-d-broad-removal/specimens/standalone-gate.txt -- the fourth; its
  serial captures were overwritten before they could be recovered, so only the
  gate transcript survives.
• docs/planning/green-program/sockets/serials/775/round3/r3-production/gate-{1..5}.txt
  -- the 5 boots of the shipped narrow arm.

Filed from the #775 round-4 slot; see the branch fix/775-dispatch-log-removal
and docs/planning/green-program/sockets/775-CENSUS-EQUIVALENCE-2026-09-04.md
for the surrounding work.

[ryanbreen]

Correction to this issue's title, from the #775 round-5 slot.

The title says "production-profile gate 4/9 red". 4 is the PASS count, not the
red count: the broad-removal arm is 5 of 9 red. The body already has it
right ("9 | 4 | 5", and "The prompt signature is 4 of 9"); the title is the one
loose place, and this comment is the correction rather than a retitle so the
original wording stays visible.

The 5 reds are two different failures, not five of this issue's signature:

how many	abort site	what it is
4	:953 x2, :942, :952	the doubled-prompt signature this issue is about
1	:1016	test "$(marker_count "$BSSHD_STARTED_LITERAL")" -eq 1 -- bsshd did not start

So: the arm is 5 of 9 red, and the prompt signature is 4 of 9.

Re-derived from the committed rows at the time of writing:

grep '| `broad-removal-22-records` |' \
  docs/planning/green-program/sockets/serials/775/round3/case-d-broad-removal/README.md

9 rows, 4 carrying a PASS count of 1 and 5 carrying 0, with abort sites
:1016, :953, :942, :952, :953.

One further note on the evidence, because round 5's review found that
directory's README overstating what is in it. Of the 22 summary rows, 4 have
committed bytes -- specimens/prod-r3head-b/gate-{3,4,5}.txt, three of them
with both serial captures, plus specimens/standalone-gate.txt with its gate
transcript only -- and the other 18 are host-recorded rows whose captures
were overwritten on the beast host before they could be committed. That is
what this issue's "Committed evidence" section already said; the README now
says it in the same words instead of "this directory is the other 22".