Native type-description name transport
Run: scratch/runs/20261121-sims-typeinfo-transport/.
The Nov 17 boundary notes describe an earlier tree. Main already contains
a3ed13388 (owned type-description/inherited IDispatch transport) and
fa85608f8 (native by-value VARIANT transport). Their tracked codec is clean;
this task neither adopts foreign WIP nor repeats registrar qualification.
The later integration note is
sims-owned-native-typeinfo-dispatch-integration-20261117.md.
The remaining ITypeInfo slot 10 was incorrectly classified as local-only. It now uses the bounded UTF-16 name-array and DISPID-array codecs, with its actual three-argument ABI. Names and parameter ordinals come from the real native ITypeInfo implementation. All returned IDs validate before publication. Wire names with interior UTF-16 NULs are rejected instead of silently becoming shorter lookup requests; counted BSTR semantics are unchanged.
The exact base contract loaded from Git fails the native regression with E_NOTIMPL while direct native lookup returns S_OK. The corrected proxy returns the native HRESULT/IDs for mixed-case, unknown, member and parameter names. Independent native memories, a nested lookup during real ReadFile suspension, cancellation of a completed lookup, malformed names/counts and a malformed late ID pass with no partial caller publication. The existing real proxy/stub OpenCAB-shaped returned-interface, descriptor release, lifetime, retirement, disconnect and allocation-failure controls also pass, ending with zero owned allocations, exports, requests, pending frames and descriptor leases. Native codec/owner/continuation regressions and the full locked build pass.
ITypeInfo Invoke still needs a design for its process-local instance pointer;
GetTypeComp, AddressOfMember, GetDllEntry and GetMops remain explicitly
unsupported by this transport. Full installation and native Quit are not
established by these controls. Any bounded original outcome is recorded in the
run's final result.json.
ABI source: ITypeInfo::GetIDsOfNames.
The single bounded original on published 63d47fe54 confirms that genuine
OpenCAB returns S_OK and ISetupCABFile calls proceed, followed by successful
component collection/item calls. This confirms the earlier integration's
progress; the name-lookup correction is qualified by its separate native tests.
The next typeinfo rejection is IID DEF5FEEC-225A-11D3-88AA-00C04F72F303.
An offline probe using the actual native SLTG metadata finds eight slots and
one rejected method: slot 5, output/retval pointer to VT_SAFEARRAY (27).
This needs an owned typed SAFEARRAY output codec and lifetime, not a validation
bypass. A later ISetupComponent slot 41 callback path also reports a stdcall
stack mismatch; its cause is not established by this run.
The personally reviewed whole 800x600 image still shows the support-file error 0x80004001. No physical input or native Quit is claimed. The parent reaches its 45-second guard and writes that image, but orderly shutdown stalls. Independent 80-second supervision sends TERM then KILL to its owned group and reports no leaked processes. The registrar's wait/both EOF receipt is complete; Embedding's wait/both EOF receipt remains incomplete, and killed-owner exit audits are unavailable. There was no control endpoint, subsequent endpoint request or replay.