▋
docs/testing/vuh2573-native-guest-magic/README.md

VUH-2573 native guest Magic evidence

PASS: 20261020-185351_net_guest_native_magic_1, product/main 9b6287b, protocol 29 unchanged. All 71 steps and three segments passed. The guest and host each select Magic with one Down, cast native Fire (10 → 0 MP), and naturally recharge to 10. The opposite native party HUD matches both changes. This supplies VUH-2568's guest native MP proof on single-PC loopback.

Captioned clip shows host left / guest right, native guest Fire followed by native host Fire, then a labeled cut to natural recharge completed. Initial RAM setup is labeled. Passing scenario report is attached with the clip to VUH-2573.

Mechanism

Magic was eligible. The guest's neutral player-class clone also ticked the canonical player's shared FIELD_COMMAND through 0x3A8980 → 0x3FD690 → 0x3B2340, consuming the same menu edge twice. The new guard skips that tick only for already-neutralized noncanonical actors with both command pointers bound to the shared controller. The canonical native path and spell eligibility remain intact; no save or MP writer is added.

Native contract and diagnostic correction records the addresses. The old spellCount and 0x3F3110 lead described the Drive list, not Magic. Probes before the fix reproduce the skips at 10 and 100 MP while the native callbacks admit Magic/Fire.

Method and limits

Two Sora profiles, one PC, loopback relay bound to localhost. The fixture uses pad input without focus changes. Initial Fire availability, 80 HP and 10 MP are RAM setup. After that, both MP costs and refills are native; no MP writes occur. Original Fire bytes are restored before closure. No in-game saves.

The 18-second presentation uses the first 14 seconds of the two 18-second renderer action clips, followed by four seconds after natural refill. No speed changes; audio omitted. All 360 presentation frames were inspected on 15 labeled sheets: game pixels and captions only, no private desktop content. The guest clip starts in the Fire submenu; one-Down selection and exact MP values come from scenario memory assertions. This is not a latency measurement. Two-PC reciprocal acceptance and other character combinations were not run.

Checks and closure

  • Release clone-neutral and native-party-HUD tests: 2/2 passed.
  • Clone-neutral Windows AddressSanitizer passed. MSVC UBSan was not run.
  • Independent inject review: ADOPT for live validation, followed by the passing run. No blocking defect found in the pinned ABI, canonical pass-through, shared-controller ownership, partial-install rollback or save safety.
  • Four protected saves, 801 source files, 14 game-root files and 527 asset files unchanged. Sealed inputs unchanged; owned games absent; relay 27813 free.

Retained bundle

  • approved-live/: report, native clips/captures and exact process logs.
  • presentation/: captioned MP4, edit commands and all-frame review sheets.
  • safety/, sealed-inputs/: closure inventories, product copies, checks, independent review and pinned source revision.
  • probe01/, probe02/, their safety folders and re/: causal controls and static traces, retained with their original diagnostic labels.
  • raw-pins.json: retained-file size/SHA-256 seal.

The lead published the qualifying files. The returned evidence manifest records retained objects and durable links. The store takes media and non-Markdown files of at least 16 KiB; Markdown and small text stay out by design. The approved-live report's 71/71 PASS and independent review's ADOPT conclusion are summarized above, and the complete report is attached to the issue.

Every returned manifest entry was checked against its PC file's size and hash. Only this README and evidence.json belong in Git. Raw media/products stay in the evidence store; smaller source records remain in the PC bundle.