Sims iKernel slot41 continuation diagnostic, 2026-11-17

Evidence: scratch/runs/20261117-sims-slot41-native-continuation-installer-progress/result.json.

No production repair is justified yet. The single authorized passive diagnostic recovered the previously UNKNOWN saved and terminal callback frames. The precise internal transfer into zero remains unresolved; no additional original was run. Full T-db74d863, BW2 and IS11 criteria remain open and unchanged.

The original uses module SHA-256 f3568ba6cd1c1982cf3c79eebada41823dd9c3d6a21cc6ac9e48491af4ed6c19 and region map SHA-256 8e15e720bcfded5902defa7e323661b824baf53ebc25de0fdcad3e8620982bf9, retained by the preceding VARIANT SAFEARRAY run. This diagnostic reuses those exact artifacts. Launch Git SHA and relevant binary diff, source hashes, media readbacks and the small passive require-hook are retained. No source archive, media copy, genuine oleaut/stdole execution or registry seed was introduced.

IID wire c34e8191f0b5d21180b900104b1f6cea, slot41, resolves to native 0x004210c4. Independent native type metadata describes HRESULT and one interface pointer, IID wire 68207eaa55cbd211809400104b1f9838. Static inspection of the pinned original iKernel (985eb55e...fedb931) finds the argument at [ebp+0xc] and the common epilogue at 0x00421afe: ret 8. This agrees with the host's two words, this and interface. It is not evidence for changing the argument count or relaxing the stack check.

The new diagnostic observes two real calls to that function. Token1018, server 0x083f9c64, argument 0x0054f4f4, returns S_OK with exact ESP 0x074ffe88, restored callee-saved registers and original FS/SEH head. Token1340, server 0x083f8534, argument 0x00595d54, fails:

Field Saved / expected Before restoration
ESP 0x074ffe88 0x074ff4d0
EBP 0x074fff6c 0x074ff504
EIP suspended bridge 0x07504c28 0
FS 0x004ac60c same
FS:[0] 0x074fffe8 0x074ff4f8
Current thunk 0x07504c28 0x07501dc8
Native private IP 0x1a1b91d8 0x1a2eafd0
Yield reason 23 0

The terminal stack is 2488 bytes below the expected caller, with callee state still live. Its last native thunk is not the successful call's final thunk. This localizes an early internal transfer/return but does not distinguish an API's wrong return frame from an exception/unwind or continuation defect. The passive observer did not retain that thunk's metadata or the final native transfer source. No process-memory recovery or further original was attempted. The old UNKNOWN ESP fields are superseded only for this new observed call; the old original's missing frames remain missing.

Before the diagnostic, actual native controls passed for correct ret 8, real error HRESULT, wrong ret 4/ret 12, an extra argument, changed callee-saved register, native trap and retirement refusal. Every applicable control checked parent restoration. The existing native yield12 suite also passed fourteen I/O/nesting/error/cancellation/retirement/output-lifetime cases, stale-token refusal and replacement-epoch non-restoration. No ESP/token/epoch/root/current owner/cancel/retire/flags/SEH check was changed. No production code was changed, so no new canonical-build or repaired-installer result is claimed.

The diagnostic's original sequence again crosses native VT12 SAFEARRAY creation and its element operations before the slot41 failure. The client receives 80010108, and the personally reviewed whole 800x600 image shows the ordinary support-file error 80004001. Those HRESULTs do not establish a failure of ordinals201/202. The installed overlay remains byte-identical; zero new installed files and no usable startup are demonstrated.

There is an explicit diagnostic restriction violation: the reused launch recipe retained the preceding run's --trace-api-mem options, producing bounded API argument memory output. The new frame observer itself records registers and SEH scalars only and preserves the native run budget and calls. A queued-cancel attempt raced with launch; the supervisor recorded SIGTERM and the one original still ran. This is not a preguest failure and does not authorize another run. No physical input or forced guest call was sent. The one selected final whole image was captured before requesting runtime shutdown.

An independent watchdog was armed before native launch. The genuine registrar exited0. The embedding child required SIGKILL; its main-thread final provider audit is absent, and guarded state merge was correctly refused. Child actual waits and both EOFs precede parent shutdown. Parent, watchdog, supervisor and outer actual waits complete; the parent exits0. This was host cancellation, not native Quit or installer completion. No box was created (ID and expiry null, state NOT_CREATED, stop receipt NOT_APPLICABLE). No workers or deployment.

Next work needs an authored native counterexample that identifies the terminal transfer at thunk 0x07501dc8 before any causal repair. The current evidence does not identify that thunk's API/continuation metadata. A further original diagnostic would require renewed owner authorization; this worker stops with the existing authorization consumed.