IW process-handle selective review — 2026-10-20

Historical source last updated, 2026-10-20. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

Disposition: aliases still absent; send transport already integrated. Do not apply either historical commit wholesale. The retained alias implementation has a sound E6 object/handle separation for the tested subset, but is not accepted as a complete process-handle implementation. No production change or main integration is made by this review.

Task: IW-DUPLICATE-PROCESS-HANDLE-20261011. Reviewed main is 2c58e9dc5d42ecc170b7c86c5b9011a1b265b62b; alias candidate is eeef32f9091b1bff7ad64d163f09a8822c0914bd; send candidate is db7a1a57e. Both historical commits exist, and neither is an ancestor of reviewed main. That ancestry fact does not mean their changes are absent. In particular, the db7a1a57e host snapshot itself has no _processAliases, _processObject or _duplicateProcessHandle: it is not the alias candidate plus a later send patch.

node docs/re-notes/iw-process-handle-selective-review-20261020.js

What is actually on main

Area Evidence and result
Alias allocation and ownership Absent. host.js has none of the three symbols above. At 2357 it still allocates child IDs from _children.size, and at 2400 its process_ctl only implements child exit-code/terminate operations 0/1. There is no E6 table, op3 duplication, ops4–8 query/rights/terminate/close/wait, stable object, restart identity, or alias teardown.
DuplicateHandle API Absent. src/09a7d-handlers-shell-file.wat:251 has no E6/process branch. Its fallback copies arg1 when host_fs_duplicate_handle returns zero. lib/filesystem.js:616 returns zero for an unknown non-file-namespace value, including pseudo-process 0xffffffff. Thus a nonnull output can still receive the contextual pseudo handle with success; no target ownership is established. This is source-path proof, not renewed guest execution.
Alias close/terminate/query Absent. No E6 cases in src/09a0-handlers-base.wat (CloseHandle:4444, TerminateProcess:6424) or src/09a0b-handlers-base-late.wat:939 (GetExitCodeProcess).
Alias waits Absent. lib/thread-manager.js has no E6 branch; E4 single wait at 1517 still infers lifetime from exit code 259.
Unsupported-host response Absent. lib/host-imports.js:3063 returns -1 for all process_ctl operations; test/run.js:4031 has only child operations 0/1. The alias candidate's explicit op3 ERROR_NOT_SUPPORTED (50) is not present.
Alias regression Absent. Main test/test-child-process-object.js remains the fe279617e child-object test; it has no DuplicateHandle export, target-owned alias, two-process alias lifetime, or alias RPC tests.
Worker send transport Present through selective integration 8cad41987, an ancestor of main. host.js:4778 installs endpoint resolution/dispatch, process send resolver, per-link wait ownership; stop at 4844 retires endpoints; the Worker loop notifies readiness. index.html:2149 includes the broker.
Foreign HWND and WAT send Present. lib/host-window.js:509 reports broker support. The foreign scalar-message block at src/09a5-handlers-window.wat:4031 is byte-identical to the block added by db7a1a57e, including error50 for unsupported routes and yield10 synchronous state.
Broker, driver and tests Present and subsequently evolved. lib/process-send-broker.js implements per-link waits/retirement; lib/guest-thread-host.js:1087 resolves sends with owned wait tokens and rethrows unexpected failures at 1128. Main includes test/test-process-send-runtime.js, broker/admission and ownership regressions. It does not merely contain the original process-wide serialization/failure-to-zero code.

The source integration is documented in docs/iw-main-integration-20261015.md and docs/iw-current-lifecycle-port-20261015.md. docs/iw-combined-native-result-20261015.md records a later full-build module 04ede400070b3acbdcc2ad5f3c033bcf051e3d2da9a4d28a449324d23ab0ba17 and eight historical native arms, with the initial fixture failure preserved. Those records are historical conformance evidence, not tests rerun here. General send-cycle redesign, ShellExecute reconciliation and native/runtime requalification are outside this task.

The TODO's older statement that both commits were absent from main0304 must not be carried forward as a current source verdict. Its current active-review status is appropriate; its completion criterion is not established on main. Likewise, older “not loaded or wired” and “held” paragraphs in the game notes describe their dated snapshots.

Retained alias correctness and defects

The following applies to eeef32f90, not main.

Correct bounded E6 behavior. host.js:4684 resolves a handle in its owner's table. _duplicateProcessHandle:4698 resolves source and destination process handles in the caller, then the source handle in the source host; allocation goes into the destination map. Equal numeric handles in different tables need not identify the same object. The alias entry holds object, access and inherit flag; it does not copy the contextual pseudo handle. process_ctl:2324 separates exit code from signaled state. Closing one E6 entry does not terminate the process or invalidate other entries. The lifetime prefix at 4735 sets exited, preserves all 32 exit-code bits, clears the departing owner's handles and releases the object's host reference. Restart at 2782 creates a new object, leaving retained aliases on the old exited one. Monotonic E6 allocation avoids stale-handle reuse and fails at exhaustion. The finally block closes E6 sources on duplication failure.

The API plumbing is explicit: E6 CloseHandle/TerminateProcess at src/09a0-handlers-base.wat:4270/6219; query rights and result storage at src/09a0b-handlers-base-late.wat:792; DuplicateHandle at src/09a7d-handlers-shell-file.wat:251; E6 single/multiple wait handling at lib/thread-manager.js:1454/1666/1684. Multiple waits validate invalid aliases before consuming another ready object. These WAT and multiple-wait paths were read, not executed by this review.

D1 — NULL output and close-only calls are rejected incorrectly. The host requires frame[16] != 0 and returns -87 without allocation. With a NULL target and CLOSE_SOURCE, it deletes the E6 source in finally but returns -6. The WAT wrapper translates that into FALSE/error6. Microsoft's DuplicateHandle contract permits NULL output with duplication and documents NULL target for close-only use. These are supported API shapes, not invalid handles. The executable characterizations reproduce both mismatches. Any correction must also avoid the wrapper's unconditional output store for those cases.

D2 — legacy E4 close lifetime is not reconciled with E6. E4 child records can be duplicated through _processHandleEntry, but ordinary CloseHandle only intercepts E6. E4 falls through the existing pipe/token/etc. checks to the filesystem; lib/filesystem.js's closeHandle returns true for unknown non-file handles. It cannot remove _children. The extracted filesystem method reports success twice for the same E4 value. Therefore ordinary E4 close leaves the record resolvable/duplicable, unlike DUPLICATE_CLOSE_SOURCE, which explicitly deletes it at host.js:4714. This is an existing E4 defect left unresolved by the candidate, not a claim that its E6 close deletes the process. CloseHandle closes a handle while other references can preserve the process object. Do not conflate deleting an E4 lookup entry with terminating the child.

D3 — E4 and E6 disagree about the same exited process. For a child exiting with code259, E6 op8 returns signaled using object.exited, but E4 op0 still returns259 and the retained ThreadManager's E4 branch at1458 returns WAIT_TIMEOUT/blocks. A fixture with one shared object and its E4 record reproduces these divergent inputs/results. E4 process_ctl termination also uses259 as its live predicate. The E6 fix does not reconcile the original CreateProcess handle. Main additionally retains the older GetExitCodeProcess E4 -1 sentinel collision: a legitimate exit0xffffffff is reported as invalid handle (src/09a0b-handlers-base-late.wat:945). The retained alias branch's broader history already separated validity; do not assume that prerequisite arrives by applying the eeef patch alone.

D4 — memory-release claim is limited to the alias object's edge. An E4 record still holds rec.wine even after object.wine becomes null. The alias object releases its host as intended, but a parent retaining its child record can still retain the child host/memory. Not a measured heap leak. (Recovered from branch task/iw-process-handle-selective-review, Oct 2026.)

Additional acceptance boundaries: OpenProcess remains a deterministic E2 handle for the current PID with the older access model; it is not a new closeable E6 entry. The E6 inherit flag is stored, but inherited-process transfer is explicitly unimplemented. Generic-rights normalization, arbitrary cross-process kernel objects and CLI IPC are not proven. A live source-table owner can duplicate an alias to an exited object; using an exited process as the table owner correctly cannot resurrect its closed table. Full teardown, asynchronous launch ordering and host restart were not exercised end-to-end. Main's stop still stops children; the held alias snapshot removed that behavior in its history. A future selective port must review that prerequisite as well.

Regression and original evidence boundaries

The retained test/test-child-process-object.js was read in full. It covers real two-WASM-memory alias transfer, actual Worker RPC, query-only rights, terminate0xffffffff, independent close, failed duplication CLOSE_SOURCE, exit259 and teardown. It does not cover D1, normal E4 close, E4/E6 wait agreement, or child-record host retention. Its original full-build dependency is why it was not rerun under this task's restrictions.

Main test/test-process-send-runtime.js was also read: its literal0xe60001 is a scalar payload for the send test, not an allocated or queried process alias. Its successful delivery therefore cannot prove alias support on main.

Original evidence read: the game notes in eeef32f90, db7a1a57e, current main, and the append-only board's 2026-10-11 10:53/12:11 and 2026-10-12 12:06 handoffs. The records say:

The exact local folders scratch/runs/20261011T1052Z-iw-process-aliases, 20261011T1154Z-process-send-runtime, 20261011T1116Z-iw-process-alias-original and 20261011T1157Z-iw-send-runtime-original are absent. No screenshots, raw timing logs or source seals from those runs were independently recovered or revalidated here. The committed handoffs establish what was recorded; they do not fill that evidence gap.