PTCT current-executor capture — 2026-11-20
Latest evidence:
scratch/runs/20261120-ptct-static-fixture-lifecycle-original-20261120/.
The untouched original now has two qualified whole current-display frames with
visible natural-clock scene change on each actual software and WebGL executor.
Both originals close through trusted ordinary Escape DOWN/UP, ExitProcess(0),
and guest/render retirement. Browser, native supervisor and independent finite
disk monitor have actual clean waits and pipe EOFs in both successful scopes.
WebGL is Chrome ANGLE/SwiftShader; no physical GPU, texture correctness or
Windows visual-parity claim follows from these captures.
The gateway exposes the binding's existing fixed-deadline PREBIND operation; failure permanently closes the gateway. Fifteen focused tests pass. The driver uses two predetermined independent capture intents, each with exactly one bounded PREBIND and one ticket. It checks cross-capture owner, executor, guest-instance and module identity. It never reopens a refused chain. A prior single-chain pair captured its first frame then correctly refused held paint; that entire runtime scope remains closed.
The fresh locked shared-tree build passes, with its owner and independent monitor joined cleanly. It validates the changed helper/tests against the shared tree; its generated module remains distinct from the immutable cc4fb822 module exercised by these runtime cases. Exact publication state and source patches are recorded in the latest run, without incorporating foreign WIP.
The authored native fixture now uses existing messageDriven/drawOnce support, without a timer. Actual software and WebGL fixture captures plus ordinary native close pass. Held-paint refusal also occurred with drawOnce, so a timer cause is not established. Actual fallback, role, module-generation, endpoint- generation, executor-context substitution and WEBGL_lose_context controls each refuse in separate fresh scopes. Shared GLTron helpers remain unchanged.
Demonstrated driver repairs: pin the registered original EXE and CRT at their actual requested URLs; wait for DOM completion before trusted pointer movement outside the output (the painted touch cursor otherwise obscures it); and wait boundedly for a real completed DRAW rather than capturing startup clear frames. The original registry, guest clocks, rendering and texture semantics are unchanged. Media stays at its existing test/binaries paths, referenced by size and SHA-256.
Two monitor lifecycle defects were reproduced and repaired in the run driver: disconnecting an already disconnected IPC channel, and treating the monitor's own planned disconnect as unexpected owner loss. Controlled regression cases also verify owner SIGTERM cleanup and actual owner/monitor joins. None explains the historical tool143 signal sender, which remains unknown. Earlier UNKNOWN, floor-cancelled and refused scopes are retained, permanently closed, with their missing individual waits explicit. Current actual floor monitoring is per-case; a root free-space check is never treated as a capacity guarantee.
The paragraphs below describe the earlier capture implementation and its historical incomplete runs. Their missing controls/close/motion are superseded only by the named fresh cases above; their failed scopes are never replayed.
Earlier evidence: scratch/runs/20261120-ptct-current-native-capture-motion-20261120/.
The predecessor /dev/shm/ptct-current-executor-20261118/ remains preserved.
The two owned implementation/test files were recovered from that predecessor.
The observer and capture gate retain the exact executor, instantiated module,
context, endpoint generation, initialized publication and ordinary chooser joins.
This follow-up adds PTCT-specific host-consent ownership/review checks because
the reused GLTron helper hard-codes its own app. Consent is an ordinary trusted
host-button click with independent bounded mouse-UP; it grants no guest chooser
command. Fourteen focused tests and five existing publication tests pass.
The final locked build after these changes also passes (build-final.log).
The actual Chrome Worker export-function/location/bytecode join now passes,
first against standard module 78214bbe…, then against module cc4fb822… in
both actual guest and render Workers of authored native x86 GL fixtures.
See complete hashes, original media references and source manifests in evidence.
WebGL here is Chrome ANGLE/SwiftShader, not a physical GPU claim.
The shared locked build passed, but its module/mirror belong to foreign shared
WIP: layout a3131d81fa59cc8a. Serving that module against immutable HEAD's
mirror (5de202609f96ea56) refused the first fixture before guest execution.
The repaired harness compiled exact Git HEAD
fd430d03e9d6037ed7054872fb5e0174526b6a6e in memory and retained its distinct
cc4fb8225e07c1136c7e200484fb321827f20a756b71b599748237786fdc692a
module. Runtime JS uses that immutable graph plus the four hash-pinned,
capture-only overlays; no foreign WIP is adopted. The shared-build pass is
not misrepresented as a build of this immutable runtime graph.
Two more measured harness problems were repaired: the ordinary fullscreen
helper rejected PTCT's app identity, and gl-renderer=software did not select
the GPU-stream software backend. The existing d3d-renderer=software selector
does; the actual-class gate correctly refused the mismatched earlier launch.
Actual software and WebGL controls retained whole current-display lossless captures of the authored yellow triangle on dark blue with black sidebars. Both were personally inspected and match the independent literal output oracle at every one of 2,019,840 pixels. Both refuse actual module-hash, endpoint-generation, role and module-generation faults. WebGL additionally refuses actual executor context/backend substitution. These are controlled authored-fixture faults, not original-game modifications. Native fallback and actual context-loss controls remain unrun; unit fallback coverage is separate.
Software's first ordinary Escape close did not produce ExitProcess within ten seconds. The next WebGL scope prepared passive input diagnostics and ordinary focus-click/independent keyboard-UP, but was externally terminated before that close phase. Its supervisor received SIGTERM at 00:31:09.107 UTC, before its 00:33:10.009 deadline, and escalated to SIGKILL. Actual driver wait is SIGKILL (137); actual browser wait is SIGKILL with both pipe EOFs. The supervisor reports no leaked processes. Signal sender/cause is unknown. No RAM cause is inferred. No endpoint request or replay follows this closed scope.
This external termination is the irreducible runtime blocker. Ordinary native close, complete controls, paired natural-clock original motion, and publication remain unproved. Code stays WIP; no commit/pull/push of incomplete qualification. All predecessor/foreign holds remain intact. No box was created: identity, expiry and stop receipt are null; state is NOT_CREATED. No VM lifecycle action, worker launch, Python, Wine, frozen stepping, private clock change or deployment. Sampled /home free space was above the floor; continuous compliance cannot be claimed (another agent reported a transient dip during the overall session).