# Session 21 verification design — close session-20's environmental PENDINGs
2026-10-22, builder crew (micro-agent circle)

## The state

Session 20 (2026-10-21, kido_sched_destroy) left its conformance review at
16 MET / 0 PARTIAL / 0 DEFERRED / **2 PENDING (environmental)**:

1. aarch64 cross-compile of the touched TUs — the cross toolchain present
   in session 19 was absent on session 20's VM; apt install was in flight
   at session end.
2. kido-reloc self-test rerun on the settled tree — destroy does not touch
   reloc, but the entry's closure required the rerun ("Update:
   cross-compile and reloc rerun before closing this entry").

Session 20's "Remains" then queued exactly one actionable verification
item for a tooled VM: "cross-compile rerun on a tooled VM." The rest is
N8toriousAi's call (D1 KIDO_ERR_NO_TASKS), Phase 4 (real BUBO peer +
bare-metal transport), or UEFI/RPi4/MMU — none of which are this
session's to guess at. The tree is byte-untouched since 2026-10-21
21:17 (kido-sched.c/.h, the destroy files) — confirmed by timestamps.

## The piece

A verification session: the full hosted regression battery on the
settled tree (same flags, same TU sets, -Wall -Wextra -Werror, zero
warnings required), the reloc rerun (gate-closed + flip-engaged,
PENDING #2's exact content), the audit fixture test, and a fresh
re-attempt of PENDING #1 with the evidence recorded either way. Zero
Gate-file modifications: pristine sources, /tmp scratch only.

## Positions

- **P1 regression keeper.** Run the whole hosted battery before any
  claim: sched (with destroy §§50–59), reloc both flavors, image, alloc,
  resonate, interop, mem, hal, entry, boot integration, audit fixtures.
  A battery that skips the modules it "didn't touch" is how fence
  breaks hide (session-5 precedent: the self-test hung on a corrupted
  free list that only a linked test exposed).
- **P2 pending closer.** The reloc rerun is not redundant with session
  20's battery — session 20 explicitly deferred it ("reloc self-test
  pending (rerun after tree settle)"). This session IS the deferred half.
  Both flavors (gate-closed 19, flip-engaged 28) on the settled tree.
- **P3 toolchain realist.** The cross-compile re-attempt must be an
  action with a recorded outcome, not a sentence that says "pending."
  `apt-get install -y gcc-aarch64-linux-gnu` was executed: "E: Unable
  to locate package" — main pocket still unreachable (the session-13
  pattern, fourth occurrence: 13/17/20/21). That IS the evidence for
  PENDING #1 staying open. Do not hand-wave; do not pretend the
  binutils (/usr/bin/aarch64-linux-gnu-*) that ARE present are a
  compiler — as/ld/readelf cannot compile kido-sched.c.
- **P4 honesty auditor.** No QEMU is owed by this session. Nothing in
  the boot path, asm, or relocation logic changed since session 19's
  QEMU run (destroy is hosted-provable C + the arch-split ctx zeroing
  that mirrors the file's existing #if pattern). Session 20's "No QEMU
  owed" stands; repeating a QEMU run on an unchanged boot path would
  be theater. The .S drift check: both .S files are byte-identical to
  session 18's object-gate proof, so re-assembling is offered as a
  drift detector only, run through aarch64-linux-gnu-as (present),
  claimed as nothing more.
- **P5 scope keeper.** D1 (KIDO_ERR_NO_TASKS) is N8toriousAi's call; Phase 4
  and UEFI/RPi4/MMU are future by design. The session delivers one
  verified piece: the settled tree is green on hosted and the two
  PENDINGs are adjudicated with evidence.

## Challenges

- **C1 — Is this session just re-running session 20's battery?**
  Answered (P2): session 20 ran the battery but recorded the reloc
  rerun as explicitly deferred. The tree then froze. This run executes
  the deferred part and re-gates everything else against the frozen
  tree. That is the closing action, not a repetition.
- **C2 — Hosted tests cannot exercise the ARM64 ctx-zeroing #if
  branches in kido_sched_destroy.** True and stated. The hosted run
  proves the destroy logic (refusal gates, free-then-clear, re-init
  round trip) on the shared C path; the arch branches are covered only
  by inspection (they mirror the file's existing #if split). This is
  precisely PENDING #1, which remains open with the apt evidence.
  The limitation is named in the ledger, not buried.
- **C3 — Should the audit fixture test run again?** Yes (P1): the
  scripts are untouched since session 18, and the fixture test is
  the only execution the gate gets without the toolchain. Cheap,
  real, pins the gate's parsing core.
- **C4 — The boot-integration link set was never recorded verbatim;
  is the chosen set authoritative?** The set is derived from the
  sources, reported in the ledger, and verified by the link itself:
  kido-boot.c + kido-sched.c + kido-alloc.c + kido-image.c +
  kido-reloc.c (boot's kido_main needs the scheduler; the scheduler
  needs alloc/image/reloc; hosted probe is a stub, so kido-mem.c and
  kido-hal-arm64.c are not linked — their self-tests cover them).
  If the link errors on an undefined symbol, the set is wrong and the
  error is the evidence; the ledger reports whichever set actually
  linked. (Verified in-run: the five-TU set linked clean.)
- **C5 — Why not assemble the .S files as the "partial gate" for
  PENDING #1?** Because assembly without the C compiler proves
  encoding only, and the encoding was proven in session 18 with a
  disassembly match. The drift-detector run (C4-adjacent, P4) is kept
  to its honest size: `as` on both .S files, reported as "no drift,"
  nothing more. Overruled as a substitute for the cross-compile.

## Convergence

The session is a verification session in the lineage of sessions
6/8/11/14/18: run what the VM can run (hosted battery + fixtures +
.S drift check), attempt what it might run (cross-compile via apt),
and record the outcome as evidence. PENDING #2 closes on a green
rerun; PENDING #1 stays open with a fourth dated data point that the
main pocket is unreachable. No source changes, no new module, no
bootable-kernel claim. The full build plan still queues for N8toriousAi;
first boot is his to witness.

## Preserved dissent

- **D1 (overruled):** defer the entire verification until the
  toolchain returns, and do one combined proof session. The
  toolchain has been absent in 4 of the last 9 sessions with no
  predictable pattern; the hosted half is real evidence and the
  reloc rerun was explicitly owed. Waiting would hold the ledger
  hostage to apt.
- **D2 (overruled):** count the apt failure as sufficient and skip
  the hosted battery too ("nothing changed since session 20's
  battery"). Timestamps are a claim, not a proof; the battery is
  the proof, and it is cheap. Run it.
