Globe native close: nested DLL unload on the owning Worker

Ordinary title close now reaches native ExitProcess(0) and guest/render retirement in both actual WebGL and software browser originals. Repair commit 1ad9afd9201b40a3985421906436bd0023c73ba4 changes only lib/guest-rpc.js and adds test/test-worker-nested-dll-unload.cjs. Evidence is scratch/runs/20261118-dx-globe-native-close-cause-repair/; its verification.json and final result.json index the actual waits and observations.

The preceding paired-rendering note already proves moving scenes. This task used its retained native destroy-handler, main-loop and cleanup disassembly first; it did not repeat a rendering campaign or contact the old UNKNOWN runtimes.

Observed boundary

WM_DESTROY enters Globe's 0x402d3b arm and cleanup 0x403841. Cleanup clears initialized (0x40e9e6), then releases six objects, and finally sets done (0x40e9e2). The first four Release return sites execute. The fifth return site 0x40390a does not. The sixth and cleanup return also remain unexecuted.

The materially extended diagnostic captures the existing synchronous-send timeout marker, without changing native CPU state:

CADE5000 00564AE4 00000005 00000002 00010001 00000000 00000000

This is native EIP 0x564ae4, yield 5, WM_DESTROY, hwnd 0x10001 at sync depth

  1. The loaded d3drm.dll base is 0x514000, original base 0x64780000; the EIP maps to original 0x647d0ae4. Selected disassembly is retained as d3drm-final-release.txt: the immediately preceding instruction at 0x647d0ade calls imported FreeLibrary, followed by ret at 0x647d0ae4.

The synchronous USER callback loop services yield 5 through host.service_load_library. The Worker previously brokered this import to the page. The page explicitly refuses to service DLL yields for a Worker, because its native instance is idle. After the existing bounded callback loop exhausts, the callback is abandoned and the caller registers restored. The native main PeekMessage loop then resumes with initialized 0 and done 0. Window disappearance alone was never evidence of successful cleanup. The 20-second original deadline remained unchanged; both diagnostic baseline scopes closed UNKNOWN permanently.

Repair and qualification

The Worker import now services only pending unloads (yield 5 and null load name) through the existing synchronous process-boot loader on the owning CPU. This preserves native detach, loader gates, return value, and dllRetired publication. Loads requiring asynchronous file I/O retain their existing path. No interpreter, renderer, worker scheduler, timeout, or held diagnostic source was changed.

Before changing a fresh original, the diagnostic was qualified with an authored real native PE in an actual browser Guest Worker: normal indirect COM Release returns, sets done and reaches PostQuitMessage/GetMessage WM_QUIT/ExitProcess(0). A second real PE stalls inside Release and qualifies the native timeout marker. The exact FreeLibrary fixture then shows baseline callback abandonment at yield 5 and, with the repair, native Release return and done 1 before native exit and retirement. Its baseline GetMessage synthesizes quit even after callback abandonment: native exit alone is therefore explicitly insufficient for that fixture. The done flag and callback-return counter distinguish old from fixed.

The tracked regression runs a real x86 WM_DESTROY in a Node Worker and checks FreeLibrary TRUE, real DllMain PROCESS_DETACH, module retirement, callback tail, CPU restoration, invalid-handle FALSE and exactly one retirement event. Against the old import with an actual page-refusal broker it fails precisely at the native callback-tail assertion after 64 refusals; with the repair it passes. Two earlier baseline harness attempts lacked that broker and timed out; they are retained as unsuccessful harness attempts, not the old-fail proof.

Worker API batching passes. The callback ownership census remains unchanged. flock /tmp/wine-build.lock bash tools/build.sh passes before the code commit. Strict pull and push both return 0. The published dd9 dual-executor regression files are byte-identical; known Tunnel ExitProcess(0) controls and their retained evidence remain intact. No unchanged Tunnel campaign was run.

Affected originals and individual waits

Both fixed originals serve the same retained native module a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8, Git runtime graph f6962ba5b1cf2829a66f5cc4d80b293c30719b37, renderer 06f27181f9c977cd1aed99270040efddca00ed75, and published RPC repair. Only the qualified bounded diagnostic and recorded observer suffixes are added in memory. HTTP hashes verify the bytes actually served. This isolates the repair from foreign work in the shared checkout and from the newer locked-build module.

After personally reviewing the whole private display, ordinary trusted mouse DOWN/UP at page (306,31) closes the title button. Both originals execute all six Release return sites and cleanup return, set done 1, clear all six object pointers, traverse native main-loop exit, and log direct native ExitProcess(0). Guest and renderer retirement resolve with no pending retirement error. The page-side guestResult stays null and is not used as exit proof. Globe uses its done flag to leave PeekMessage; no PostQuitMessage/GetMessage claim is made for this original (those are covered by the qualified fixture).

Each original has actual separate driver/supervisor, Chrome, Xvfb and openbox wait receipts: code 0, null signal, both output EOFs. Native supervision reports no leaked children, no deadline termination, and all tracked PIDs are absent. Browser tmpfs, display socket, lock and authentication file are removed after retrieving the unique evidence. The 90-second runtime, 20-second close bound, independent 105-second supervisor plus 15-second cleanup bound remain finite. No forced termination counts as native success.

Minimum sampled free /home bytes are 1,556,611,072 (WebGL) and 1,554,309,120 (software). These are 100ms samples, not a continuous-disk claim. An earlier fixture launch was refused before runtime when disk was below the floor. Only owned closed text was losslessly compressed with verified round trips.

The first software whole-display capture preceded its visible composition; a fresh whole-display capture in the same still-open finite scope was reviewed before close. No artificial warmup, debugger busy attestation, deadline extension, or old endpoint was used. Other early fixture harness errors (title overwritten, post-retirement memory observer, unquoted fixture mode) are recorded separately and do not contribute to the pass.

Recovery uses Git commits, per-scope source manifests, relevant binary diffs, recorded observer suffixes, diagnostic/config, and owned new-file patches. Game media and the retained module are referenced by path and hash, never copied. No source tree snapshots, Python, Wine reference runs, public deployment or box were used. Box ID/state/expiry/stop receipt: null / NOT_CREATED / null / null. Foreign WIP and every security hold, including Blood timer work, were preserved.