IW current-lifecycle integration checkpoint

Base: eb26d3a567d0a71ba6c1672afbebd82836d36d65, fetched main on October 15. The earlier held e2d10ebb chain forked at 4fde32fa; it does not contain the current loader lock or generation-aware unload implementation. Its note about absent shared readiness describes that old branch, not current main.

The first checkpoint selectively ports the ten transport/frame/wait/generic yield commits through 357727bab. Main's native loader, reference counts, generation pins, callback gate and terminal thread-exit behavior are preserved. Broker cycle admission is unchanged. The loader commit e2d10ebb is excluded: its conflicts in dll-loader.js and process-boot.js need semantic migration. This is held integration work, not a main merge or game compatibility claim.

Two integration corrections were necessary even though the patches applied: the old set_last_error export duplicates the existing export in 10-helpers.wat and is removed; the scheduler fixture now requires ExitThread intent to remain unsignaled until the Worker reports completed detach/exit, matching current main.

Passing source checks: actual Worker/local scoped generic yields, wait reservations, frame ownership, wait identity, both wait repro tools, broker ownership, scheduler (56 checks), browser cache graph, test-tier membership, and diff whitespace. These use inspected tiny Wasm fixtures or JS mocks.

An accidental invocation of test-process-send-runtime.js attempted a local full source compile, contrary to the source-only constraint. Validation failed with duplicate export set_last_error, before instantiation or runtime. The duplicate was removed; the native suite was not rerun. No generated build file or ignored artifact appeared in this worktree; disk after the attempt was 2.4 GiB available. No Boat, browser, guest application or remote runtime ran.

Next: adapt scoped loading to main's beginDllLoad/endDllLoad generation pins, nonblocking callback gate, finishDllInitialization/rollback contracts and native completion transitions. Preserve the exact frame/epoch checks and separate host publication receipt. Native two-Worker lifecycle and original-game validation remain queued for an explicitly granted runtime slot.

Loader migration boundaries found after the checkpoint

e2d10ebb was attempted on the isolated port and then aborted; none of that loader patch is present in this checkpoint. Conflicts are not merely export lists. Current main already owns the relevant native mechanisms:

• loadDll holds DLL_LOAD_LOCK before staging, uses an instance-owned row index, retains ROOTS or PINS on reuse, and waits for INITIALIZING by generation. loadedModuleNamesLocked reports only LIVE rows. The earlier audit of absent lock/count reservation was accurate only for the obsolete branch.
• beginDllLoad pins the captured provider generations across byte lookup; endDllLoad releases them. Capturing a bare set of names loses that ownership.
• tryDllCallbackGate is the nonblocking admission primitive. The async helper is not appropriate inside a synchronous scoped endpoint command. Busy must return pending without consuming the yield; successful admission must remain owned across host awaits, nested callback yields and initializer steps.
• finishDllMain calls native complete_dll_initialize for the original row generation. finishDllInitialization releases the mapping pin only when appropriate. The old scoped initializer invokes neither: pasting it in would leave a successful image INITIALIZING forever. FALSE, fault and terminal exit have distinct attached/failed/quarantine semantics that must survive the port.
• Yield 5 also represents FreeLibrary, FreeLibraryAndExitThread and deferred unload draining. The native take_dll_unload_kind is destructive. Its result must be captured once in the frame/epoch transaction before generic load-name lookup. A retry cannot take it again or turn an unload into LoadLibrary.
• Current stop/retirement drops local frames and terminates Workers. A newly retained callback gate or provider snapshot cannot simply live in a closure that disappears on termination. Actual retirement needs an ownership-checked release/abandon protocol that leaves incomplete image pins quarantined and cannot release a replacement thread's gate. This requires native/host identity review, not an unconditional atomic store from the host.
• Completion must preserve main's generation-bearing mapped/retired resource events (registerMappedDll, isCurrentDllMapping) as well as the held chain's separate host-publication receipt. The receipt alone is not native readiness.

This is a prerequisite port, not an implementation of these loader steps. Yields 3/5 remain outside the owned service on this branch, as on 357727bab. No callback-loader success, cross-thread loader fix or runtime compatibility is claimed. Existing source tests cannot discharge native lifecycle requirements.

Prepared native readiness conformance

test/test-dll-same-module-concurrent.js now holds A before its actual x86 DllMain until B has inspected the INITIALIZING row. It removes the 300 ms scheduling assumption and the misleading host flag set before DllMain ran. Both independent native Workers must see the name absent while initializing and present after TRUE; B's returned mapping must actually be LIVE in the native lifecycle table. Same HMODULE, one mapping, same-owner reentry and native guest LoadLibrary parking checks remain. This fixture has only syntax/diff checks on the new branch; execution is queued. A worker failure must fail the suite, not be treated as initialization success.

After an explicit sole-runtime grant and a matching canonical build, run:

node test/test-dll-same-module-concurrent.js
node test/test-dll-concurrent-mapping.js
node test/test-dll-lifecycle-races.js
node test/test-dll-worker-message-lifecycle.js
node test/test-dll-import-graph-lifecycle.js
node test/test-dll-terminal-lifecycle.js
node test/test-process-send-runtime.js

Those existing ordinary-loader tests are prerequisites. The next scoped port also needs actual Worker/local tests with the native lifecycle ABI for two owners racing one name, delayed lookup plus provider unload, same-thread nested load/unload, FALSE/trap/terminal exit, lost mapping/publication replies and owner retirement. Keep frame/epoch and generation checks simultaneously; tiny mapping fixtures without the lifecycle ABI cannot establish this.