Independent legacy caps bootstrap review — 2026-10-20

Reviewed 7b6335dd8fb144d9533da603f2e42540db6a3d84 (origin/task/legacy-caps-early-exit-oct20). Accept the fixture correction; no blocking defect or production correction found. This review performs only offline verification of saved execution evidence, with no build or runtime repeat and no main integration.

The actual diff adds one functional call, e.init_dx_com_thunks(), immediately after successful load_pe, plus read-only terminal/object observations and a report. The PE generator, oracle and synthetic controls are byte-identical. Canonical source 1262cbfcf8b5881543f2fb18c0f80b4bd9429306 calls the same helper after PE loading in host.js:3787 and test/run.js:5482. The fixture requires the export directly, appropriate for its pinned module; normal launchers guard the export for compatibility.

The generated initializer (src/09b2-dispatch-table.generated.wat:22509) constructs real method tables, beginning with IDirectDraw's 23 slots at API 978. init_com_vtable (src/09a8-handlers-directx.wat:1591) allocates tables, writes dispatch thunks and updates the shared registry. DirectDrawCreate passes that initialized global to dx_create_com_obj, which writes lpVtbl. This is ordinary runtime bootstrap, not fixture-specific result injection: it does not write expected HRESULTs, descriptors, callback counts or oracle records. Its heap allocation also explains the changed entry stack address.

The baseline observation has DirectDrawCreate success and object 0x08221000 but lpVtbl/slot20 zero, terminal EIP zero, incomplete guest and untouched record-four poison. Stack return address 0x403087 and the saved x86 call [eax+0x50] at 0x403081 support the null-target diagnosis. Poison is not an API HRESULT. The corrected observation has vtable 0x41986c, slot20 0x75001d8, completion 1 and terminal ESP baseEsp+4. The original oracle independently replays successfully on all 47 call records, 45 descriptors, 55 guard pairs and callback counts 3/3/2/1.

Evidence identity:

scratch/runs/20261020-legacy-caps-bootstrap-review/{review.js,tests.json} preserves the independent replay recipe and receipt. It checks 35 existing artifact/reference hashes (including actual saved module bytes), candidate bytes against Git, all 6,441 staged-source OIDs against the canonical tree, 26 recorded host dependencies against canonical Git bytes, observation addresses/metadata against the unchanged manifest, syntax, scoped diff, and 1,053 synthetic negative controls. All pass. The source manifest is a staged subset of 6,898 Git entries; the 457 omitted entries include assets and notes, but no src/, lib/ or host.js. An initial overly broad reviewer assertion that the staged manifest equaled the entire tree was corrected accordingly. Historical pre/postflight receipts report unchanged remote sources and 29,620 prior files; this review does not independently rescan that remote filesystem. Recorded execution is Node v24.18.1, one diagnostic and one corrected run.

Qualification is a real x86 public COM API fixture executed in the full pinned emulator module with production host imports and renderer:null. It covers public discovery/creation/QI, Device1/2/3 GetCaps at 172/204/252 bytes in aligned and unaligned storage, fixed 236-byte Device7 descriptors, selected line and triangle texture/filter/address masks, and live EnumDevices descriptor copies. It is not a direct WAT helper-only test or Windows-native execution.

Remaining coverage gaps:

• Legacy 3/3/2 callbacks reach final HAL; cancellation there is indistinguishable from exhaustion. D3D7's first-of-two cancellation is discriminating.
• Only selected descriptor fields are asserted, not every captured word. Enumeration canaries protect copied buffers, not internal driver buffers; earlier software-device descriptors are not qualified by the final HAL copy.
• Release execution/stack cleanup is checked, not exact reference counts or complete lifetime correctness. Nonvolatile register preservation is not independently asserted by this oracle.
• No invalid-size/null/error matrix, cube allocation rejection, texture-format enumeration, mip storage, sampling, rendering, backend parity or gameplay qualification. Other modules/engines/browser workers remain unqualified.
• No screenshot, per-instruction trace or engine-native disassembly exists. The diagnostic uses block-boundary observations plus guest bytes/stack.

Reproduce only the offline review with node scratch/runs/20261020-legacy-caps-bootstrap-review/review.js. Its referenced evidence is scratch/runs/20261020-legacy-caps-early-exit; that directory's RECOVERY.md pins reused module/helper/source dependencies in the preceding native-run bundle. Keep these references together; no large artifact was copied. Historical runtime commands are not rerun authorization.