MCM owned storage and Escape publication boundary
Evidence: scratch/runs/20261120-mcm-owned-storage-publication-quit-repair/.
The two requested parent results, appnotes, exact launch scripts, retained
output ledgers, trace and frozen runtime source graph were read. No production
code was changed and no owned storage culprit was established.
Final metadata for the quarry parent totals 7,446,528 allocated bytes across
135 paths; the aborted diagnostic totals 516,096 bytes across 53 paths. Both
launch scripts already put Chrome profile and TMPDIR under /dev/shm.
Known direct driver/Chrome FD metadata was read only after checking their
recorded process start tokens; no FD contents or platform process inventory
was inspected. These observations cannot establish historical allocation
peaks or attribute a global /home free-space drop. External storage
attribution remains UNKNOWN. No speculative placement or logging change was
made.
One fresh diagnostic used the exact 2,420,767-byte f5ae1ac84e3d5e13bb093aca53b8aa2fe18c000cc891587a5d75b46eb559b2ba
module, the prior coherent in-memory Git/diff reconstruction and all 311
verified existing media references. Full page frames were inspected through
the ordinary menus and quarry. There is no new throttle/turn claim; the
owner-verified parent driving evidence is reused unchanged.
Escape reproduced the actual publication refusal. In the same failing read,
the presenter layer and window existed, _gpuFrameLayer still equaled that
layer, and instance/process ownership matched. However, the layer was stale,
_dxFrameLayer was a different object, and its write sequence was 182544
against the GPS layer's 182540. GPS reported 994 presents and stamp 30781.
The replacement canvas exposed no _waD3dimVersion. The screen included in
that same exception still shows quarry terrain and the rider; it does not
prove a published menu or a complete native replacement frame.
The frozen source explains a possible CPU handoff: the D3DIM GPS layer's
stale predicate checks native present ownership/CPU sequence, and the native
DIB presentation path can replace a GPS _dxFrameLayer. The actual failed
read proves the stale/replacement predicates, not which native call or
publication-join dependency caused them. No claim of a proved native producer
defect follows from source reachability. A repair requires an actual
same-owner complete replacement publication and its native producer/binding
provenance, followed by old-fail/new-pass and replacement/retirement negatives.
That proof is absent here. Relaxing the GPS predicate would be an unproved
fallback, so no sampler change or fresh ordinary Quit was attempted.
The first UNKNOWN permanently closed the diagnostic. There were no later endpoint requests or replay. Independent Escape UP ran at 101 ms, but the failed hold is not a success. Direct Chrome and board-child waits have both pipe EOF receipts. A run-local allocation-observer timer mistakenly kept the already-cleaned driver alive; its in-memory allocation samples were not recovered. The exact driver was terminated, its parent actually joined it with both pipe EOFs, and exit 143 is preserved as cancellation, not success. Allocation peaks therefore remain UNKNOWN. Production custody code was not changed. Both native receipts report no leaks; owned profile/temp are absent.
After runtime cleanup and direct joins completed, the separate metadata
monitor observed /home free space fall to 1,483,743,232 bytes at
1795165724099. That permanently closed task runtime eligibility too. There
were no later endpoint requests or fresh scopes. The global drop still has
UNKNOWN writer attribution; later recovery does not reopen the task. The
monitor's actualRuntimeLaunched:false describes that metadata helper only,
not the whole task; the diagnostic launch receipts remain authoritative.
Two command review descriptions used anticipated selection names while the retained frames still showed track selection/loading. The images are the evidence; no separate reviewed bike-selection/start-screen claim is made. No retained source/media copies, workers, Python, Wine, shared Q2/WAT edits, or deployment occurred. Localhost box ID is null, state NOT_CREATED, expiry null, stop receipt null; root retains shutdown ownership. This task ends at the precise missing replacement-publication proof, with ordinary Quit unproved.