D3D9 GetBackBuffer selectors and resource ownership

Reproduced and repaired on both shipped modules on 2026-11-06. Ordinary generated x86 PE calls through public device, swap-chain and surface COM vtables demonstrated 34 unexpected successful invalid requests: 17 per module. Every unexpected success returned the current implicit buffer's identity and acquired a reference that made the following Reset fail with D3DERR_INVALIDCALL. Public Release of that unexpected successful acquisition restored Reset eligibility.

The device route ignored chain, buffer and type selectors. The swap-chain route already rejected nonzero buffer indices but ignored type. The repair validates the device selectors and swap-chain type before wrapper lookup or AddRef. Successful calls retain the existing implicit-buffer lookup and resource lifetime helpers. The production change is confined to src/09ad-handlers-d3d9.wat; no resource helper, renderer or creation policy changed. Existing valid-call guest memory accesses and unimplemented fail-fast handlers retain their paths.

The public topology was established before applying validation: GetNumberOfSwapChains returned one, GetSwapChain(0) succeeded, and GetPresentParameters reported one actual back buffer for requested creation counts 0, 1, 2 and 3. Both getter routes agreed on valid buffer-zero identity. For those configurations, device buffer-one requests previously returned the same buffer-zero identity while swap-chain buffer-one requests failed. Multiple-buffer presentation remains unsupported; accepted larger creation requests continue reporting one actual buffer. This repair introduces no new creation rejection or additional-chain implementation.

The contract is documented by Microsoft's device GetBackBuffer, swap-chain GetBackBuffer, back-buffer type, and presentation parameters references. Buffer indices are bounded by the actual count, MONO is the valid D3D9 type, and successful acquisition requires Release. The cited getter contracts do not specify failure output contents. The regression records those contents for diagnosis and never asserts that failure clears or preserves the pointer; it never releases an output from a failed request.

test/fixtures/d3d9-reset-fixture.js now emits a 197120-byte PE32 with 111202 code bytes and 487 statically checked COM call sites. Public slots and argument counts are checked against the production API table before Chrome starts. Every executed COM call verifies ESP and EBX/ESI/EDI/EBP preservation, with distinct nonzero initial register seeds. The existing raw-code and incorrect GetBackBuffer-slot controls remain rejecting preflight checks.

node test/test-d3d9-public-reset-web.js executes the same generated PE in the production owning CPU Worker and shared WebGL Worker on standard and compat modules. Each module records 25 invalid selector cases, including out-of-range and large unsigned chain/buffer indices, LEFT/RIGHT and an unknown type, device/swap-chain parity, and the actual-count boundary for all four requested creation counts. The baseline completed its guest assertions and cleanup on both modules, then correctly failed the selector oracle. The fixed run returns D3DERR_INVALIDCALL for all 50 requests, followed by 50 successful Resets without releasing failure outputs.

Public controls also verify subsequent valid identity, description and owner; two successful acquisitions plus explicit AddRef/Release block Reset until the last external surface reference is released; a successful surface keeps its released parent device usable through GetDevice, Clear and two literal LockRect pixel witnesses; and final surface Release retires that parent. GetRenderTarget verifies an independent RT0 binding before and after implicit buffer lookups. A peer device preserves all 64 literal pixels through the selector sequence. The existing 516 safe-pixel comparisons across 14 rendered readbacks, healthy Reset, strict executor refusal, failed Reset, supported recreation and peer isolation still pass. Baseline and candidate successful frames and contact-sheet PNGs are byte-identical; both sheets were personally reviewed. These are generated regression controls and previews. Original-game behavior, historical27 completion and broad crash freedom remain unproved.

Validation passed:

Evidence: scratch/runs/20261106-d3d9-backbuffer-selectors/. Baseline and candidate source records pin Git commit 05ee8c9851f9aaf9d98b15179f47af01be6a3fbf, relevant binary diffs, new-file patches, source hashes and retained distinct module bytes. The candidate standard module is 41d3ed6a2b52b46d25550dd54f582c574a7acc48f2024c3e01f878bdafbe470f; compat is 2c94cba0b6138e51d2d69ed19f938f2cb6abd7b83cd3cc0363ce232119cbaf7a. Actual browser-loaded hashes matched those files. baseline-summary.json, comparison.json, both browser result files, command logs and the generated PE retain the executed reproduction and repair evidence. No source archives, game-media copies, original-game execution, boxes or deployment were used.

The ordered Conan0350, KKnD0400 and AoW0410 final releases and their resource cleanup were independently checked before shared edits, builds or execution. gate.json fixes the local execution window at 05:29:17.917742–06:14:17.917742 UTC. All test execution finished at 05:32:22 UTC. Final source identity and owned-process/port/profile/lock cleanup records accompany the result, which is published last. No test boxes were assigned; box ID/state/expiry are N/A.