Original Delphi 5 evaluation: host transport boundary

2026-11-16. Installation, credential requirements, IDE execution, samples and normal native Quit remain unproved. The documented root INSTALL.EXE executed and created a window titled Delphi Trial, then the host desktop transport failed with EBUSY before any input was sent. No screenshot survived the exception. Window-creation logs do not establish a personally reviewed visible installer scene or installation success.

Evidence: scratch/runs/20261116-delphi5-original-trial-install-ide/result.json. The coordinator's 09:35:13Z instruction explicitly ended original testing for this task after the transport repair; no original was launched after that instruction. The failed endpoint was never rebound or sent input.

Original route and exact inputs

The complete disc was extracted once to the owned canonical path test/binaries/candidates/delphi5-60day-eval/media/. Its 38 files remain untouched; no cabinet payload was unpacked and launched separately. The ISO is 81,330,176 bytes, SHA-256 bcd14e38635f4b3b4c7656f1ce884c26ba60224d7876681a7e3b7382e5c38b0e. Recovery and original-media provenance are in delphi5-evaluation.md.

The CLI mounted the disc as read-only D:, started D:\INSTALL.EXE with D:\ as its working directory, used the default Windows 98 identity and the real clock, and enabled ordinary process spawning and guest threads. No serial/password, trial date, expiry code, activation, guest application bytes or provider implementation was changed. guest-overlay/ and guest-registry.json were configured below the canonical candidate directory; neither was written before the host exception. There is no installed state to resume or export.

The actual runtime used Git commit 3bcb7bef7e388ff469cdaa7bb297ecc7a76d3bad and module SHA-256 ef7de08a1a1ea564ea7ff2fbd1e3dbbd56cbfc07c19d3db7fcb45c9bac4b5fb8. Runtime JS/JSON and the module were held in sealed memory descriptors. The borrowed GTA2 module holder was read-only; this task held and closed its own descriptor and never stopped that holder. runtime-inputs.json records the individual source hashes. Git plus the recorded helper patch provide source recovery; there is no source archive or persisted WASM dump.

An earlier host-only preparation failed waiting for desktop authority startup before guest initialization. Its records are retained in startup-prerequisite-v1/. Narrowing the sealed source view to the runtime dependency closure removed that preparation failure. This was followed by one actual root-installer execution; no input gesture was attempted in either execution.

Transport cause and repair

The desktop completion function writes ready=1 and then notifies its caller. The caller can consume the response and reset the same buffer to ready=0 before that notification happens. The next wait then wakes on the preceding notification. Previously receive() treated any notification as completion, returned the old response, and allowed another request while the bridge was still handling the current request.

The original trace records main-return for poll 7344 before the bridge even received poll 7344. Poll 7345 then failed with EBUSY. A real Node Worker fixture independently reproduced the stale response before the fix.

lib/process-desktop.js now rechecks the ready flag after waking. One monotonic deadline covers the entire wait, retaining the existing 3000 ms timeout. Repeated notifications cannot extend it. The focused Worker tests pass for the current response, repeated unready notifications and sealed source loading through a real Worker. The existing desktop transport observation regression also passes, including actual sockets, lost replies, authority death and terminal no-retry behavior. Its first invocation used a relative preload path and failed in the test harness; that failure and the corrected absolute-path run are both retained.

The required flock /tmp/wine-build.lock bash tools/build.sh passed with an actual exit-0 wait. This build qualifies the shared checkout for publication; its output was never used to rerun the original. The separately recorded foreign working-tree changes were not included in this worker's commit.

The original also logged COMCTL32 requesting HeapCreate(0x04000000,1,0). The known HEAP_SHARED implementation boundary remains unchanged. There was no MakeCriticalSectionGlobal trap or demonstrated cross-process heap use in this execution; the first terminal failure was the host transport error. See the Sims native boundary.

Lifetime and next test

The 1200 s native deadline and independent 1220 s watchdog were armed before launch, with a 180 s Quit/Cancel reserve. Native PID 3335288 exited 1 after approximately 12 s including startup; its supervisor actually waited and closed both output streams. Watchdog PID 3335263 observed that wait and exited 0 without firing. No installer child was created, and no native Quit was claimed. The owned descriptors, PIDs and port 45651 were closed. No box was created, so box ID, expiry and stop receipt are inapplicable.

A fresh, separately authorized original test should use the published desktop repair, capture and personally review the complete launcher scene, then follow only its ordinary installation UI. It must establish the actual serial/trial password prompt if one occurs. The prepared controller and later launcher edits were not exercised on the original and remain uncommitted WIP; they are not qualification evidence.