I Always Picked Rogue as My Helper. She Took Hulk's Chair.

I always take Rogue as the helper.

When I pick two characters in Marvel vs. Capcom, she is the one who comes in to save me. Other people spam Colossus and the assists that actually annoy you. I found out in certain tournaments that Rogue ranked lower, so you could take a higher-tier point character and still score. I kept picking her. She looked cool. Nobody used her. So I kept using her further and further.

That is the whole reason this ROM exists as an idea. I have been playing with her as a helper for so long that I wanted to play as her.

Jin's chair looked like the right place, because Storm had just fitted it. The probe said Rogue does not fit. The chair that actually opens is the one the Magneto costing reserved and never spent: Hulk $06 and Orange Hulk $26, a 100% tile-share pair. Evicting one frees zero. Evicting both frees bank 2.

Exact V18b status — 2026-08-25

The current acceptance candidate is private V18b SHA-256 fe6b487b7f75fe7b562e549892e872a492d92060001956fe0ef34b099195cdc5. Its fail-closed ledger fixes the denominator at 97 requirements: 17 ACCEPTED, 1 WAIVED, 63 OPEN, 15 HOLD, and 1 FAIL. That leaves 79 blockers. The project does not convert infrastructure rows into a completion percentage, so parity remains null, playability remains false, and a parity release remains unauthorized.

The previous 30 August V18b playtest package is no longer offered. A new one-file builder is in validation, and its download is temporarily withheld. That distribution change closes no ledger row and does not promote V29, parity, or full playability.

What is accepted is deliberately narrow:

• walk and idle body-push: 24/24 across both Rogue owners, sides, and facings;
• stolen Hadouken metadata/hit/block/chip/debit: 24/24, plus 8/8 combo cases and 8/8 negatives;
• Goodnight pose and captured-stun scope: 8/8;
• bounded point/bench placement for Hadouken and Goodnight: 8/8.

The clip that made Ryu look as if he were simply jumping beside Rogue came from the quarantined failed V13 archive, but the visual report remains open rather than being dismissed with that provenance. Exact V18b's eight-lane capture is byte-exact to the accepted private-V16 regression; it is not a fresh XvSF source comparison. The earlier source matrix reached only 5/8 exact 93-sample pose windows, 7/8 exact held-stun windows, and 0/8 source-exact complete release trajectories. A new same-MAME source/target gate must bind action, phase, relative XY, velocity, unique victim OBJ ownership, and the composited PNG on the same frame before the “jumping beside Rogue” observation can close. The ledger therefore accepts only the narrow historical pose/stun regression. Full victim choreography, natural command access, damage, presentation, release, and cleanup remain unaccepted.

That does not accept the complete moveset. CATCH 07 is the one exact ROM FAIL: source release is +140, while V18b releases at +136/+137. The sealed all-phase evidence observes damage, positional displacement, and both-fighter idle in all 32 positive Ryu lanes with zero watchdogs. That is a narrow functional result, not the full slam gate: damage RNG was unset, and seeded damage, victim stun/pose, ground contact, cleanup, a fresh throw-idle gate, extended play, and negative lanes are absent. Selector, geometry, and launch remain HOLD; slam, damage, stun, and recovery remain OPEN. Exact V18b already contains record-1 param $0100; the old $0100 watchdog belongs to an earlier isolated child. Independent review receipt SHA-256: f8f1b2ddc7b35b04c1cd28774cd4efe962c1975c02dedaa609c3f4b193109cac. An earlier scheduler attempt produced no evidence: its staged XMvSF key was wrong, so MAME never left startup screens, and that capture root remains retired. R6 derives the correct 20-member source archive with key SHA-1 65905400f4462c175baa93b43e015e8596def31b and preserves all 19 non-key members; that repaired staging only. The later reviewed V27 diagnostic records ordered exact-V29 opcode reads/prefetch at $E62 and $CEA1A, but records $CA34=0 and $D970=0. V28 advances the observer into an ordinary-input Rogue-versus-Ryu match and records both addresses. Matrix M1 then varies owner, identity, facing, CATCH, and assist-attempt lanes and broadens the strict read/prefetch subset by one caller. Neither package proves instruction retirement or behavior, and no Power Drain scheduler lane or gameplay case is accepted. The source command census also proves Rogue has no persistent flight mode; the real remaining movement scope is jumps, super jump, air dash, dash links, and complete air attacks.

Exact V18b ordinary-input replay — 2026-08-30

This exact-V18b ordinary-input move replay stayed on the published candidate bytes, SHA-256 fe6b487b7f75fe7b562e549892e872a492d92060001956fe0ef34b099195cdc5. It did not test private V29 and did not build another ROM.

The signed-side FBNeo diagnostic completed 26 unique captures: 13 native Rogue lanes for public $06 and the same 13 for secret $26, with ordinary two-player controls and passive Ryu $12. Setup X was $0100; every lane kept its required positive signed distance (+64, +72, or +240), with zero side flips, zero out-of-bounds targets, exact owner identity, and intact fighter pointers.

• 8 PASS: forward air dash, the no-back-air-dash negative, airborne PP as attack-not-dash, and back+PP as attack-not-dash, for both Rogue owners.
• 18 HOLD: ground/air/rising Repeating Punch, Power Dive, and Diving Kick contact lanes. Their state/action, child, cleanup, identity, pointer, and signed-geometry checks all pass. Natural one-run damage remains non-gating until a matched-RNG or predeclared distribution gate exists.
• 0 deterministic move-contract failures, 0 invalid, 0 missing. That is the result of the move-specific signed scorer, not a claim that all generic gate receipts pass. The generic receipts also require move-specific QSound proof, which this diagnostic explicitly excludes.

Receipt SHA-256 is 76b7e4ced057f66bfd5d1739da27770874af77c7609c4461511a320ea5deab5d; capture-manifest SHA-256 is d073ebc7cb5a178c8ae681024f895d54536ac4c2301109608b2abc47f9b75c9f. An independent clean-room rescore reproduced the canonical JSON and Markdown byte for byte.

The 18 damage HOLDs now have a sealed source-only ordinal-barrier design. It fixes 18 comparisons and 54 future serial captures: 27 calibration cases at seed $1357, followed by 27 confirmation cases at $2468. Same-seed-only screenshots do not qualify—the older ground-LP trace reached damage at source ordinals 27/28, 42/43, 57/58 but target ordinals 34/35, 49/50, 64/65. Future evidence must align the complete contiguous RNG prefix, transaction, and global ordinals before comparing damage. After the single logged pre-input seed write, the barrier may not call, advance, freeze, force, rewrite, or reseed RNG. Seven exact-V18b runtime row bindings are still missing, so the package has no runner, commands, runtime root, scoring authority, or credit. Its static verifier and all 15 hostile tests pass. Package aggregate SHA-256 is da3855a65360b1d65586c3b549757b3f0039f1c91234e2cab5e5878aad9822be; receipt SHA-256 is 8a73a80c8b441e87e8833c3e1d25313487a06f23a654de61d9b11c46aa754140.

The separately sealed five-route public-Rogue/P1 trial is intentionally stricter about its one sampled damage sequence. Standing kick, crouching punch, and jumping punch pass. Standing punch and the six-button standing chain fail only that single-run equality check:

stand P      stock 4 / 8 / 10      V18b 3 / 7 / 10
stand magic  stock 4 / 3 / 6 / 5 / 7 / 5
             V18b 3 / 3 / 6 / 5 / 7 / 5

All five routes match their input schedules, exact states, contacts, cancel continuity, transition timing, and recovery. Because RNG is not controlled in this run, the two damage differences are RNG-unresolved single-run mismatches, not confirmed deterministic regressions. The receipt is SHA-256 fa73a02742d46798d817c49a77769327d91ed1d6e4f0b14f97ff571a09720989; terminal seal SHA-256 is 020c9ddf5f7fb2b57f0efdf38bb14d1c0d1bbc5fadfdf1e70a8ff26adbfd3c15.

A signed-facing expansion defines the full 40-lane chain denominator: two Rogue owners × P1/P2 × right/left × five routes. Static V2 fixes that Cartesian product, but it does not authorize execution. Activation V5 was rejected for one pre-guard import P0 and four provenance/tool-path P1 defects. V6 repaired those, then a concurrent root replacement exposed one remaining check/resolve TOCTOU P1.

Additive V7 is now sealed source-only. Its pinned verifier and 47/47 hostile tests pass, and an independent review of the exact sealed subject reports P0=0/P1=0. Retained no-follow directory descriptors and device/inode/type/mode continuity checks statically close the V6 defect, but the policy remains default-deny with no commands. All 40 bindings are structurally expressible and remain 0 executed / 0 scored. Only the five public-$06/P1/right routes have direct stock-source scope; the other 35 remain blockers, and no symmetry artifact is admitted. V7 final-seal SHA-256 is cdab8deea738e1aca7bdfcb2412b192bd9f7ce0fd66f04e86a44c26dd9fffdf8; independent-review receipt SHA-256 is cda54142b65ece9c9d3e160290c9b8073a540079d0c6e6269899d7add0647a3b. No candidate, ROM, FBNeo, emulator, capture, scoring, or gameplay was authorized by V7, so every V2–V7 chain-preparation credit remains zero. The shared behavior reporter was separately corrected so a legacy unsigned capture may prove exact absolute geometry without silently claiming signed-side geometry; the full harness self-test passes after that regression was added.

The newest gates stopped bad evidence before it became a patch

V28 records the scheduler and join

Power Drain V24 left all 52 direct-call rows UNKNOWN. V25 then mapped the candidate control surface, but its opcode reconstruction omitted MAME's inclusive word at byte $100000. V26 supersedes V25 and its review for whole-opcode-view authority. Corrected V29 OPCODE SHA-256 38171f8b3b451ce1594a436969d0998c82317a6da0f4a0f8992454f026580d23 changes only $100000..$100001, from DATA 0800 to OPCODE 6B41. The two affected all-even candidates begin at $0FFFFE and $100000.

V26's regenerated tables have zero classified-row delta: 88 direct controls, 2,740 indirect-control candidates, 6,800 return candidates, a 620-node / 685-edge root projection with 49 barriers, a 303 / 411 assumed-return scheduler projection, a strict 140 / 155 wrapper graph, and 52 still-UNKNOWN caller obligations. Receipt SHA-256: c866a565df0cc2caa683de4003b3c8ce0bfd1fb8e1117e5240746e46940b2fe8. MAME's exact-tag sources show the separate program/opcode maps and inclusive decrypt range in cps2.cpp, cps2crypt.cpp, and the 68000 reset path in m68000-sdf.cpp.

V27 supplies the first repeatable read-only startup observation. Two isolated MAME 0.287 launches produced byte-identical deterministic opcode-read cores, recording ordered opcode reads/prefetch at reset $E2, TRAP-0, $362, $38E, $E62, $105A, and $CEA1A. The descriptor-write tap also records $FF0040 with state $0C and entry $E62. Both runs also observe DATA/OPCODE 0800/6B41 at $100000. The taps can include prefetch—they are not instruction-retirement hooks—and the capped 64/256 counts are lower bounds. Most importantly, both captures record zero $CA34 and zero $D970 observations. Repeatability receipt SHA-256: 7058371a66dbbd1537e1cf8f5ae51fb406f04d9555d6300c8b3738ecd19b2322. Independent narrow-approval SHA-256: 389dd71c38c91ca54ea4f692ca0a4adcd42f9019790abbdec957427fdf96c8ee. V27 seals no Power Drain transfer and grants zero gameplay or release credit.

V28 then changes the evidence scenario, not the ROM. Two fresh isolated MAME 0.287 runs use only ordinary input to naturally select P1 Rogue $06 against P2 Ryu $12, enter match phase $0800, and complete the same neutral window. Their deterministic evidence cores are byte-identical. Each records 733 reads/prefetches at scheduler $CA34, 8,798 reads/prefetches across the 52 known caller sites, and 4,003 entry reads/prefetches at shared join $D970. Sixteen of the 52 caller sites appear somewhere in the full capture.

The verifier recomputes 1,688 full narrow, phase-gated correlations across exactly eight role_global caller sites and the four fighter roots $FF3000, $FF3400, $FF3800, and $FF3C00. The body callback exposes a guarded active user-stack top equal to the expected return PC. That is still an opcode read/prefetch witness: it is not instruction retirement, untapped adjacency, or full control-flow closure. The other 44/52 caller obligations have no full witness here, and complete joint-state, exception, interrupt, retry, and return-epoch domains remain UNKNOWN.

Repeatability receipt SHA-256: 480a0a779ac6d28c2fea2871c3717e27a59a3c576baa0d96cd7b4eff1ee03b57. After two verifier fixes requested during independent review, the retained package passes all 11 fail-closed adversarial tests and receives the bounded verdict APPROVE_NARROW. Independent review SHA-256: 318db14af26dd587b1501fd15df2b322b9acccb26044445f1f7bab3821335063. It still seals no Power Drain transfer and grants zero gameplay, ledger, parity, build, release, or download credit.

Matrix M1 broadens the observed subset once

Matrix M1 is an append-only evidence package over the same private ROM V29; it is not another ROM version. Eleven fresh initial executions cover the neutral baseline, four CATCH configurations, four MP+MK partner-call attempts, and a non-fight CSS negative. Only productive discovery lanes are repeated, leaving 13 retained captures.

The two neutral B0 runs reproduce V28 exactly: 1,688 full narrow correlations, eight caller sites times 211, with byte-identical deterministic cores. Under M1's stricter post-orientation boundary, 1,680 rows remain, eight sites times 210. That hard regression had to pass before any discovery row could count.

Only C1 and C3 broaden the strict subset. Each byte-identical repeat records 32 rows at fixed-slot caller $CB50 while the same row also has scheduler A5 $FF8000, fixed A6 $FF3400, Rogue's active CATCH $20/$32, and the victim's captured $22|$1C / $34|$36 state. C0 and C2 add no site. The four partner-call lanes each retain 36 rows during the exact eight MP+MK press frames, but add no caller and explicitly leave assist activation unproved. The non-fight N0 lane records zero correlations.

The strict union is therefore 9/52, with $CB50 the only repeatable new site and 43 caller obligations still unobserved. Broader post-attempt and retry sightings remain context only. Retry pairing is non-consuming opcode-read/ prefetch chronology, not one-to-one execution or instruction retirement.

Matrix receipt SHA-256: cf4c4c65a197edee4a849039f48064e5dc4b345d5e9c861ea98429c9dd8784da. Repeatability SHA-256: e7a5acb2ea6ccb27bef89b984a9b87b559ae7bde1981769c9c9af26d002a6163. The pinned verifier passes before and after all 16 adversarial tests. Independent review approves only this narrow design/result boundary, SHA-256 7e4be0a355dedd01c04474ba0ecfa9c8299a0cee3abbfa965cd85f49b7b90876. Every gameplay, ledger, parity, release, and download credit remains zero.

Matrix M2 tests tag-once and closes nothing

Matrix M2 completed a bounded 12-capture ordinary-input tag experiment. The baseline reproduced 1,688 V28-full rows in two byte-identical runs, and both sealed $CB50 regression lanes retained 32 rows. Across all eight tag variants, one four-frame HP+HK chord and natural release were observed, but the lifecycle passed through a both-active interval instead of the required adjacent point-only-to-bench-only transition. No contiguous 60-frame settlement was observed.

Every tag lane therefore remains conservatively classified HOLD, not a successful tag discovery. No new caller was qualified: the observed union remains 9/52, with 43 obligations unobserved. The result grants no gameplay, ledger, parity, release, or download credit, so private V29 remains unpromoted and no Rogue patch is currently offered while the one-file builder is validated.

Matrix receipt SHA-256: ed768342c393543d674b8a627b391b509f64442230516eccc4cade18010d70fe. Frozen aggregate SHA-256: 328023088328a9c36e09d9ba3a60f52dd2166939deab8b0372ae5df4fa626ceb. Independent review SHA-256: dc4eaa8b2748ec86a994c764a3be1e101699540e28785f119a282329d5e49a5f. The terminal verifier passes over 197 frozen files, and all 35 hostile tests pass. Terminal seal SHA-256: b093e339a1c5b8e8f04e2f51d4b41d4fe5ec6dd19cbe34a7cc84cf61caeb2248.

V29 restores one helper call — no gameplay credit

The newest linear artifact is private V29 SHA-256 573b7224c2160bc830d53ab5fba55389825727b5ff1ed81828983687333bf14b, MD5 3fdcf6a54b973974d4e5b61f19101cd3. It is a direct child of exact V28, not of the accepted V18b baseline. The exact-V28 helper census accounts for all 38 Rogue-blob calls to eight requested XMvSF helper families. Thirty-two still ended in the fail-closed RTS family; one of those sites had a complete native MvC homolog and an exact one-site recipe.

V29 applies only that recipe:

$39C19C  4eb90039cf8c  JSR $39CF8C  ; graft RTS stub
        →4eb900014ef2  JSR $014EF2  ; native MvC inner reset

The six-byte request changes three logical DATA/OPCODE bytes and three raw mvc.10 bytes. It adds no wrapper and consumes no cave. Every other ROM member is byte-identical to V28; ZIP order and semantic metadata are preserved. The builder check and 10/10 focused tests pass, including a fresh byte-identical serialization. Build receipt SHA-256: 6c422a59b81bb3b793cc2113ac118e1d9d94c55c069f253eb427d245d8163db5.

This is the late D4=$08 global-reset homolog reached from Rogue's e6=$2E terminal family. The V29 build receipt establishes that call boundary only; at build time no emulator or guard was launched. The later V27 opcode-read/prefetch trace supplied no evidence for this helper call. V28 records a narrow scheduler/ join subset but does not identify this helper call as an instruction-retirement event or test its behavior; no clip was recorded, and no ledger row changed. Walk/body-push should not be patched from the old video report: the exact V18b 24/24 gate is accepted and its program surfaces remain byte-identical through V28; V29 does not touch them. CATCH 07 still releases four or three frames early, Power Drain still lacks a complete source-family contract, and Goodnight victim choreography remains broader than its accepted pose window.

The prepared exact-candidate denominator is 2,956 ordered cases across 58 groups: 210 XMvSF source cases and 2,746 target cases, covering all 85 runtime- testable ledger lanes. They remain 2,956 planned / 0 executed. The late D4=$08 source/current/patched trace is an additional zero-case scoring obligation over existing complete-special lanes, not a synthetic gameplay PASS. A distinct sibling review independently reconstructs that matrix and passes 30/30 adversarial tests; receipt SHA-256 83a9f1d92db59520c429aadd0cbf0f2605cef5fd3951cae7bd18243e208418b9. That approval permits static catalog adoption only and grants no runtime or gameplay credit.

The class-$48 cleanup lock remains separate. A later carrierless-acquisition audit shows that delete-before-set can close same-address ABA only if every nonzero or unknown writer of the four fighter-slot +$267 bytes passes through one wrapper. That semantic writer denominator is not closed, and a byte-exact fail-closed wrapper proof would still follow it. The native-move promotion matrix therefore remains hard-locked and zero-credit pending those ownership proofs, exact request/authority/capture binding, scored MAME execution, and a new ledger revision. Neither V27 nor V28 satisfies that gameplay matrix. The current audit receipt is 13fa0827324c60f4be9cccb1a38734e300866a32f6ea7cfb7d12be06e8014272.

The next Power Drain helper audit also stops before a ROM. It proves the candidate-only one-site shape $39AE3E: JSR $39CF8C -> JSR $3AB22A, three operand-byte deltas and no new cave, and closes 4/8 mandatory static gates. But the inherited subtype-$08 evidence matches timing and expanded-cell counts 12/12 while raw lists/attributes match 0/12 and transforms only 1/12. That is not visible animation parity. The next blocker is a semantic map of the 12 source records through list parts, attributes, transforms, tiles, palettes, and rendered geometry; all-context owner roles, every-exit cleanup, and independent review remain open. The 22/22-tested audit therefore grants no postimage, build, runtime, or ledger credit.

The append-only geometry pass now resolves that surface question without authorizing the patch. Native MvC $181BF6 repacks all 12 source bank-3 records into target bank 0: 12/12 oriented-cell sets, 303/303 cells, 12/12 alpha planes, 12/12 pen planes, and palette selector $0A->$10. Comparing against target bank 3 fails all 12. The defect has narrowed to exact V29's $3AB18A: MOVE.B #$06,+$47(A6); an eventual bounded repair would change only the exact-site byte at $3AB18D from $06 to native $00. The preimage occurs twice, so this cannot be a blind pattern replacement. Independent review, live palette/RGB and cadence, cleanup, owner/player/facing coverage, and stock class-$38 regressions remain open. The 24/24-tested receipt grants no ROM, build, runtime, ledger, or parity credit.

Native-move R5 is now statically sealed for 572 fresh MAME processes: 156 stock-XMvSF source point cases, 312 exact-V18b point cases, and 104 contextual bench cases. Its 42/42 Python tests, 5/5 Lua syntax checks, bundle recomposition, and sealed R2/R4 lineage pass. Execution is intentionally locked and has run zero cases, so it grants no gameplay, percentage, ZIP, or release credit. Preparation receipt SHA-256: 2ff608d42961e81626d01918cf3e66a773010a372c9fa847f563a6d0aea02377.

R6 adds a hard-locked execution contract for only a two-case Repeating Punch smoke—one source case and its matching public-V18b case. Its audit and 15/15 tests pass, but every runtime binding is null and execute exits 3. It cannot launch, score gameplay, attribute a ROM failure, or authorize later phases. Receipt SHA-256: c521c20462d1cb0baf2cfe29b3401353dfaae099fcdecbad52d7846ef90b0c48.

The parallel R6 pointer preparation is still HOLD_POINTER_NORMALIZATION_UNBOUND_ROM_NOT_EVALUATED: 0/9 R5 pointer fields is reviewed or statically complete. It also exposed a contract conflict: R6's proposed entries shape and EXPLICITLY_REVIEWED_FOR_PHASE_A status do not satisfy R5's required fields/roles shape and REVIEWED_EXACT_RELOCATION_MAP_BOUND status. R6 demands that reviewed map before its smoke, while the map itself needs fresh observations first. Receipt SHA-256: 1d7e39c61a229dbf98e05859ef1f5e22ea7186fbbfa1c2b20bee854727fc71b6.

R7 resolves that circularity only on paper. Its P0 bootstrap preserves the exact 572-case R5 order to record nine raw pointer fields without comparison, PASS/FAIL, ROM evaluation, or scoring. The static audit and 19/19 tests pass; all 41 runtime/review/scoring bindings remain null. No emulator, capture, map, runtime case, ledger row, or gameplay credit exists. Any eventual scored run must recapture all 572 cases after independent map review. Receipt SHA-256: b271c657a69272f3a86e1132e3983e21a06407c16442cb2213a7a6bcb765c1a8.

Later runtime-adapter work remains locked too. R8c closes three earlier static defects, but an independent 10/10 adversarial review reproduces six activation blockers: an authority-bypassing public spawn path, an asynchronous-exception gap around spawn, request-controlled non-global locks, unsafe executable paths, identity drift during the 572-case loop, and non-durable process quarantine. R8c has executed zero emulator/guard/ROM cases and cannot be activated. Independent rejection receipt SHA-256: 29da3282926370d9fd4242f6fa3cf27d60e57e4de9e3292370e9d097cb6d272f.

R8d is rejected too. Its production entry is hard-locked, but an independently reproduced 14/14 adversarial suite reaches all 16 provider methods through a public synthetic lifecycle path that never checks that lock. Provider identity, global lock/nonce state, verified executable objects, fresh identity checks, process-tree absence, and durable keeper transfer remain unbound assertions. R8d stays unactivated and zero-runtime. Rejection receipt SHA-256: b8393d85e0c27ffa67f6e8027855b3ccdd0e38eb1624dff34147cf61cb65a517.

R8e closes that public bypass only by deleting the runtime-adapter surface. The whole-package AST audit finds no provider/backend, callable injection, predecessor import, process/lock/filesystem service, or activation path. Its zero-argument front door consists only of return 3, with all 18 future runtime bindings null. Agent and independent root replays pass the audit and 16/16 adversarial tests. R8e is therefore a static tombstone, not an executable successor, and still produces zero MAME/guard/ROM/runtime/gameplay credit. Receipt SHA-256: cea0f0a499a944fa2b2d24470d56c4e3c9a157040ca8165627825123c0b6a8b4.

The exact helper-route census accounts for all 38 relocated calls: 34 still target the CLEAN RTS stub and four are promoted. Seven families retain RTS residue. The $011600->$014EF2 homolog now has enough static ownership proof to authorize exactly one private experiment at V18b's relocated $39C19C call—not an integration, acceptance, or release claim.

That experiment is now built as quarantined private V19. It changes only mvc.10, and only three logical data/opcode bytes at $39C19F/$39C1A0/$39C1A1, redirecting $39C19C from the CLEAN RTS stub to $014EF2. Independent --check, 9/9 focused tests, and archive integrity checks pass. Its ZIP is SHA-256 31f17a70c4d00fbc0abf7dbd62997329634152ef21bd3e861cd473b4dca01c92, MD5 be79abcc37985acc705ec849e9f946bf. No emulator was run: every matched trace, owner/side, watchdog/throw, move, team/assist, and stock-caller gate is pending, so V19 earns zero gameplay or parity credit and cannot be promoted. Exact V18b remains the current 17/97 ledger candidate.

The newer private V20 archive is a separate direct child of exact V18b, not a replacement for that baseline and not stacked on V19. It installs identity-matched Goodnight victim geometry for Wolverine, Rogue, Gambit, Chun-Li/Shadow Lady, and Zangief while deliberately preserving the existing Ryu $12 route. Its deterministic static review passes; its SHA-256 is f327cf7218e5994e0730c1d2b4465084960bbbbb15ee33ba770cb8d8ecc689a6. It has never run and earns zero gameplay or ledger credit. The first V20 correlated trace/scorer proposal is also hard-locked: independent 12/12 tests show that stale sessions, reversed/interleaved events, wrong tap/PC pairs, and a self-consistent changed manifest can pass. V2 rejection receipt SHA-256: 41fca6836ea21471d5be7f94d7d6fa33c6bfdec2b7657060be7235796bc9b5c6.

The append-only V3 replacement now passes independent static review. It begins only at the exact captured-victim leaf $01418E, requires one contiguous same-frame path through $39F9E0, $3AB000, optional fallback $014236, order-only $014254, and exact A3 checkpoint $014284, and validates every complete repeated session. Stale/interleaved streams, identity substitution, and wrong-but-even fallback bases fail closed. The frozen surface is 64 mapped target lanes, 128 exact-base fallback lanes, and 24 unscored source lanes. Root replay passes 12 generator, 20 scorer, and 10 independent tests plus Lua syntax. This proves selector route/A3 arithmetic only; command, contact, choreography, gameplay, and ledger credit remain zero. Independent review receipt SHA-256: df2cd63b6a079316d6f881c8b74fae21ece521307638b8d160d34bb558943bec.

The reported Power Drain heart now has an exact-V18b/V20 static transfer. The source Rogue palette, decoded visible heart tiles, list transforms, normalized animation, motion table, constructor, and complete class-$84 producer surface are preserved. That authorizes no color patch; a matched exact-build live matrix is still required. The separate late class-$48 helper retains older defects but was not the visible kiss particle. Static receipt SHA-256: adfc2f0074939918a0386c7e6b2580a4a622a42fa8b5c0e6a7193ae72fe6cf64. An independent 9/9 adversarial replay confirms the source-exact transfer and the no-color-patch boundary; review receipt SHA-256: c4597771aa294d32da8c0f29810d702088c1980a6a7bf9a7e73238fa42d7d736. Heart presentation remains OPEN and latch/steal remains HOLD.

Private V21 is a static-only child of exact V20, SHA-256 d5d0cb249afba1c295009cd14214b920e935d4436b64997169180b7f18a2df0e. Across the same 28 class-$38 callsites, its route census changes from 25 RTS / 2 native / 1 Goodnight to 24 RTS / 2 native / 1 Goodnight / 1 Repeating Punch. The new child carries target-private $5250 (RP) routing metadata. Outer router $3AB156 sends that tagged subtype-$08 object to translated handler $3AB176 and tail-jumps every non-RP object to the unchanged $39E456 Goodnight router. $3AB176 maps only source $066134..$06617B; $06617C begins the separate class-$3C constructor and was deliberately excluded. Only mvc.10 and mvce.03a differ from V20. Build receipt SHA-256: b7d6f5f13794f91deef70e06bd54d20e71da8ba3b257ec8ff8ccd107307682d7. Maintained 11/11 and independent 14/14 adversarial tests pass with verdict APPROVE_PRIVATE_STATIC_ONLY; independent review SHA-256: e3ca0a6d5338657066a817a60687dc59d7502bc834973465f16aeb1928ade652. This is not a runtime result. All eight live gates remain, it grants no gameplay, parity, or ledger credit, and exact V18b remains the 17/97 baseline with 79 blockers.

Private V22 is the first exact current-QSound descendant of V21. Its SHA-256 is ac41a2abaec73362a27eb28d4000bcf2b6627555ad3db5caca08d02188b8502c. It changes the opcode hook, one 128-byte action row, the owner transport, and the four QSound members—nothing else. The static package has 35 source actions, 31 row actions, four common gameplay actions, 28 PCM clips, and all samples contained within their $8000 windows. Maintained 7/7 and independent 14/14 tests pass. That is static packaging, not proof that the voices sound normal. Natural owner posts, pitch, rate, loops, duration, missing/chipmunk audio, stock interference, and extended play still need exact-hash recordings.

The next defect was not another missing move. V21 had introduced two target-private pool-C tags at object +$6E: Goodnight $4753 and Repeating Punch $5250. Both MvC allocators clear only $00..$5F, and native free $0044E6 never clears +$6E. A reused slot could therefore carry the wrong identity into the next object.

Private V23 implements a statically verified patch for that narrow lifecycle surface. Its SHA-256 is 7f7f9f4f865fbb89dee2c373c095ce3083b2619479c6f672a7cae0dc44aec69f. Only mvc.10 changes, with 89 logical bytes. RP and Goodnight clear +$6E before native free. The RP landing class-$3C call uses a private clone of the already-correct native constructor and clears inherited metadata only after a successful allocation; the failure displacement changes $34->$38 so failed A0 is never touched. The shared constructor, handler, vector, and V22 audio remain exact. Maintained 7/7 and independent 14/14 tests pass. This still earns zero gameplay credit.

The specific report that Ryu looks like he is merely jumping upward with Rogue was checked separately. Exact V23 still contains the accepted captured action- $34 implementation: all six keepalive-3 sites, the class-$10 route, Ryu's $12->$39FA00->$014254 selector path, the animation/list regions, and all eight GFX members are statically preserved. That does not transfer the old video evidence. V20's selector and V23 have never been run, and V23 changes the final Goodnight cleanup route, so a fresh exact-V23 capture is still mandatory; roster-wide victim choreography remains OPEN.

The V23 lifecycle plan inherits V21's 120 RP rows once and adds 424 exact cleanup/reuse/failure/negative rows, for 544 total. Its checker, 12 maintained tests, and 18 independent adversarial tests pass. Every row is unexecuted; recording clearance and request binding are null, authorized commands are empty, and activation is HARD_LOCKED.

The animation census also corrected the broadest visual assumption. Public $06 and secret $26 each expose 278 addressable actions through the same physical tree: 1,488 records, 238 jumps, 784 actor lists, and 14,359 mapped tiles. All 14,359 tile payloads are byte-exact and the reachable tree contains zero unexpected Hulk-residue tiles. The closed mismatch set is seven timing records and six actor-list records. One empty EFFECT-00 stub explains all 19 unreachable Rogue tiles. Nine named hypers—111 records, 74 lists, and 2,100 tiles—have no static timing or actor-semantic mismatch. Public palette data is 96/96 words exact; Shadow Rogue intentionally differs 48/96. The independently reviewed core program projection is 126,704 bytes across 20 intervals.

That does not prove the supers render correctly in play. Effect ownership, background/cut-in routing, palette selection, cleanup, and retained-roster negatives remain live gates, so no graphics patch was issued. A first visual preparation was rejected because its broader surface omitted 328 core cut-in bytes. Append-only V2 restores those bytes and proves a 127,282-byte, 17-interval superset while leaving the hard-locked plan unchanged: 2,628 HOLD rows, including 2,308 developer-menu/injection rows that can never earn parity credit and 320 natural or natural-route-pending rows. No command, screenshot, capture, or runtime result exists.

Audio and pacing therefore remain live blockers. Exact V18b still has its mixed QSound row, while V22/V23 are quarantined static children. The ordinary owner-post design defines 216 future processes but has run zero. The split- clock speed design has also run zero; each owner still needs 120 cumulative within-gameplay-segment host-monotonic seconds and 240 samples in a throttled, audio-enabled same-hash run.

Why Jin looked right

Storm fitted Jin's freed bank-3 space: 12,329 new slots against 12,873 freed, +544. That measurement is why Jin's chair looked like the right place for an X-Man.

Rogue is not Storm. Same method, Captain America kept in the survivor set, zero reliance on "0xFF means free" as ownership:

‹›

matched idlegate strip (XvSF pens)

Cap stays. Cap's exclusive GFX lives in bank 0. Jin's lives in bank 3. A fighter addresses exactly one bank. Early Storm/Magneto playtests showed leftover identity: Magneto's cut-in still Captain America, Storm's HUD still printed JIN. A tile budget that "fits" while the HUD still says Jin is not a replacement. Magneto and Storm later closed two natural P1/LP identity paths on a separate candidate; Storm's gameplay HUD still says JIN. H-bank expansion of a clean Euro parent was authorized as a fallback while the Jin ledger stayed red. It is not the product. expand_vanilla_hbank_authorized is false on this path.

The chair that actually opens

The unified Magneto article already measured this family: Hulk $06 and Orange Hulk $26 share 100% of their tiles. Deleting Hulk alone frees literally zero. Deleting both frees 16,427 exclusive slots in bank 2.

Magneto did not spend that opening. He sat on Captain America $04 with MAG_TREE in the expansion band. Dual Rogue spends it.

Select graph is 15 public ids. $26 is the existing Orange Hulk secret box, not a second grid cell. Hulk's public cell is row 3 col 2, palette $02. Parent is vanilla Euro mvsc md5 e7dbef6dd363c11fddd12f0d446f647b. Source is xmvsfj.zip md5 3d424538843cfce0e03b96d73277f60b. The historical early-candidate archive—not V18b and not currently offered—was named mvsc_rogue_hulk_v1.zip, md5 a6c411ce0e9c0f35ac41a1b03bdb1fe7, SHA-256 4651010a2af0b5163e43dcb10f40d91ceff416a92761c3bec14c1f2ab7daee9f.

What the live stills actually show

‹›

idle body in the squaresmiling bust $2A576A

Shadow Rogue is the existing secret boxPattern $0C868 from Chun-Li sets bit 2 of $FF4199. One Up from Ryu $12 reaches $26. Confirm with Up+LP on that cell so $00A4CE runs for the right id.CONFIRMED IN GAME

cursor id	$26
unlock	bit 2 of $FF4199 = $04
confirm	Up+LP on the extra cell
pal src	$1C9164 · word0 $0248
MAME	secret.log RESULT=PASS

Do not poke $FF3052 after versus start. Repeating Start after the lock burns the $26 match and lands in attract with a public team. $26 is not a second grid cell. The 100% pass still includes the secret-select large figure behind her.

MAME mame_rogue_hulk_live.lua case=secret · snapshot 0001.png

The match body is RoguePalette directory $1C8344, 48 words, mismatches=0. Pal word0 is not vanilla Hulk $06F0.CONFIRMED IN GAME

HUD word	ROGUE
body pal src	$1C8344
live dest	$914000 · 48/48
68K	CLEAN $399816
secret pal	$1C9164 → $914000 · $0248

Partner in the public capture is Ryu. P2 is Strider / Gambit. Shadow Rogue versus palram is a Shadow Lady role remap, not a darkened XvSF row and not a RAM pin. Lifebar face $6EC6 has a 1px transparent frame so the blue rails and the nameplate E-ticks survive.

MAME case=public snapshot 0001.png · public BODY_PAL

The lifebar face is insetTile $6EC6. Row 0, row 31, and the right two columns are pen 15. Cap’s E keeps its trailing outline column.CONFIRMED IN GAME

paint	$6EC6 2×2
safe frame	row0 / row31 / col30–31
name	ROGUE · Cap E +1px outline
table	opcode $6EC6, data-view stays vanilla

The old idle crop filled the 32×32 and sat on the blue rails. Vanilla Hulk already left row 31 empty.

MAME case=public snapshot 0001.png crop

HIT EDITER already draws her treeDest tree $3CD180. accept_label on the dump profile is still HULK; the HIT line prints ROGUE.CONFIRMED IN GAME

editor line	03 ROGUE · STAND FOOTWORK
CG_Ptr	$003CD41A
dest tree	$3CD180 public / $3D34D2 shadow
join	196/196 names and slots
boxes / moveattr	seated; $26 aliases $06

Walk/jump/gravity rows from XvSF Rogue are seated. T85AC left at Hulk $00040002 (no XvSF homolog). CATCH EDITER is the throw oracle; her throw handlers are the next 100% slice after Power Drain.

destination_hit_deep/g000_s000.png · ROGUE_L1_TO_MVC.json

The secret match body is Shadow Lady’s morphDirectory $1C9164. Live palram word0 $0248, skin $0510, navy on gold and green pens. Not a dim Dual Rogue.CONFIRMED IN GAME

live word0	$0248
skin	$0510 mahogany
P navy	$0248 / $0127 / $0004
K cloth in ROM	$0A47 magenta / $0713
select row 19	$1D2CE4 · $0248

Role remap from stock Shadow Lady $1C92E4. Public Dual Rogue $1C8344 stays undarkened XvSF. Do not globally morph P2 portrait pal $1D4F04.

MAME case=secret snapshot 0003.png · BODY_PAL mismatches=0

Shadow Rogue is Shadow Lady, not a dim Rogue

The first secret seating darkened XvSF Rogue in place. Word0 landed at $07A9. That read as a tint, not as a character. Stock Marvel vs. Capcom already owns the pair that should sit here: Shadow Lady’s two palettes on $1C92E4 — navy on P, magenta on K, mahogany skin on both.

$1C9164 is a role remap of that row onto Rogue’s pens. Dual Rogue’s layout is not Chun-Li’s. Pen 0 is the white streak. Pens 1–5 are peach skin. Pens 6–8 are gold hair, gloves, and boots. Pens 9–10 are the khaki jacket. Pens 11–13 are the green leotard. Pen 14 is lips.

Putting Shadow Lady’s gold $0FF0 on Dual Rogue’s gold pens painted a yellow bodysuit. Navy belongs on the gold and the green: $0248 / $0127 / $0115 / $0004. Skin stays mahogany $0510. The streak follows Shadow Lady’s white→navy. Lips stay burgundy $0810. K is $0A47 / $0713. Live secret palram matches that row: rom0=$248 live0=$248 mismatches=0. Public Dual Rogue $1C8344 stays the undarkened XvSF row. Unlock still uses LP, so the live secret match shows the P colors.

A palette-only remap cannot split hair from skin: both still share pens 1–5. If the 100% pass wants a white streak on mahogany without painting the hair navy-adjacent, that is a later GFX seat, not a second pal row.

‹›

public $06 · $1C8344secret $26 · $1C9164

The yellow smear was Hulk EFFECT 00

The first Dual Rogue seating still ran Hulk’s hit spawn for L1 group 24: a 24-smear EFFECT list that stuck in the stage sky as a yellow blob (live idle window y 67–92, x 3–83). That is not Rogue art. stub_l1_spritelists() now points group 24 at an empty mode-4 sprite list at $28A3F6 on both public and shadow trees (67 sites each). After the stub the same idle window has 0 yellow pixels. The stub stays until her own dash-trail EFFECT is seated — part of the same 100% pass as her 68K.

The super window is a mask, not a rectangle

An earlier seating pointed Hulk's hyper at XvSF KAO faces ($2A75DA) — the 144×64 select heads from $0FB432. Those are not in-match cut-ins. Stock MvC hypers at $0C15EA are 176×224, origin (0,0), with a thick jagged black left edge. Ryu's is the reference. Painting Hulk's shared hyper tiles in place is forbidden: 44 opaque tiles belong to other characters.

The donor bust is XvSF table $0802C4 index 2: Rogue cel $2A6A24 (208×240), P1 pal $11A45C, P2 pal $11B07C. P1 and P2 are palette rows, not two cels. Cover-scale ate her fists and hair. compose_mvc_cutin.py now contain-scales the full opaque bust into the interior box past the p90 jagged tooth, left-aligns, and inks a 4px black stroke only on the jagged left edge. Shared Hulk hyper tiles stay untouched. The result is a new 1×1 cel on sole-owner tiles in blank GFX $1A000–$1C000.

Hulk's hyper record $0C15FA now points at $28BE38. P2 is $0C1A1A. VS KAO records stay KAO. Live MAME fills fraction $FF3273=$FF and stock $FF3274=$03, then QCF+PP. Stock drops from $03 to $00. The cut-in cel is hers. The super kit is not. Live play still crashes, glitches, and can become Ryu.

Win portraits are a different table ($080224) and a different pose. Rogue's smile, hand-out / hip cel is $2A576A with pal $11BFDC (word0 $0DD6). Storm's $11BF3C is the wrong row. That bust is blit-centered onto Hulk's win canvas as cel $28C276, then hung on $0C14D2 / $0C1572 / extra family $0C18F2.

‹›

Ryu stock hyperRogue $28BE38

‹›

Hulk winRogue $28C276

SPECIAL 00 was the first live special

Hulk's first special is Gamma Charge: charge back, forward + punch. HIT EDITER group Char Gp $2C sub $00 still prints SPECIAL 00. The animation pointer is Rogue's dash at $3D0636. That seating was her body on his charge recognizer: the first proof that a special slot on this chair could play her art. The yellow charge smear that used to trail her is Hulk EFFECT 00; that list is now an empty sprite list.

The continued 100% pass does not stop at that dash. Her CLEAN 68K now owns a detector. Power Drain is live on the same chair.

SPECIAL 00 on the dest treeChar Gp $2C / sub $00. accept_label HULK. The HIT line prints ROGUE. Now_Cnt 00/13.CONFIRMED IN GAME

editor line	03 ROGUE · SPECIAL 00
CG_Ptr	$003D0636
Char Gp	$2C $00
tag	g005_s000
68K now	CLEAN $399816

SPECIAL 00 was the first live special: her dash on the charge slot. Power Drain is the next measured special on the same chair.

destination_hit_deep/g005_s000.png · COMPACT_NAME_TO_CHAIN.json

Her 68K sits in CLEAN

The freeze still holds: vanilla Euro mvsc only, no XvSF 68K copied onto Hulk’s window at $02F6FC. That window stays parked so versus construction at $030E28 still has a legal family successor. CODE_DIR for Hulk and Orange Hulk still points at $03050E / $030516. Init at $00D14C still runs $00A4B0, which is how $00A4CE uploads the body palette. Her fight 68K lives somewhere else.

CPS-2's opcode and data views differ through the encrypted word beginning at $100000. Exact V29 reads DATA 0800 and OPCODE 6B41 at that boundary; data == opcode begins at $100002. In Hulk's first-megabyte window the views agree on about four words in a thousand, so a Rogue overlay that looks right as opcode can still read vanilla Hulk as data. Rogue's blob at $399816 is safely in the equal-view region, and its PC-relative tables read what the detector expects.

The blob lands at $399816. Reloc delta $372B82. The detector at $39C6F0 copies XvSF $026508–$026C94 — normals through Power Drain — including the stolen-id dispatcher at $026780. That dispatcher is move.w $342(a6) into a per-victim QCF, not a meter table. Ryu indexes $08. Strider indexes $0E. Copy it through so QCF after the kiss latches the move she actually stole.

Tail used 2,616 bytes with 904 of headroom. Receipt status STATIC_PAL_SAFE_GRAFT_CLEAN. Matcher $013C30 keeps the vanilla family successor $019710. Eligibility $012450 keeps $017486. Magneto’s fight 68K lives at MAG_TREE $588000 in the expansion band. Those are two different answers to “where does the guest’s brain go.” Dual Rogue does not take Magneto’s expansion path. She takes the empty CLEAN that the first megabyte would have lied about.

Power Drain is live

Power Drain is QCB+LK. The live question is not “does a special fire.” SPECIAL 00 already proved a slot can play her art. The 100% question is whether her kiss holds the person she kissed, writes the steal, and then lets her fire their QCF.

The grab veneer at $39D0CA is 94 bytes. QCB can turn Dual Rogue away from the victim, so the veneer compares $c against the opponent and writes facing into $4b and $b3 (0 = right, 1 = left). It snaps the victim $24 pixels in front of that facing, JSR’s Rogue’s hook $39BCE0 ($02915E), then JSR’s MvC attach $0187EE so the pair take actions $32 / $34. Opponent $a8 is pushed, the attach runs, $a8 is restored. Stock $0187CA is the same attach plus move.w $52, $a8(opponent). That extra write is how a defender’s X leaves the kiss and lands on the far side of the screen. The veneer is the attach I want, without that write.

Live Dual Rogue $06 vs Ryu $12, public yellow and green, zip md5 a6c411ce0e9c0f35ac41a1b03bdb1fe7:

• age 16: still walking in, face=$01, p2x=132
• age 20: action=$32 group=$20 p2act=$34 face=$00 p1x=79 p2x=115 steal=$0012
• age 204: stolen Hadouken latch special=$08 / group $2C sub $18

Thirty-six pixels. Face flipped toward Ryu. Steal id $12. That is her kiss, then her stolen QCF. The detector did what $026780 is for.

The next 100% slice on this same object is the projectile. MvC type $48 is Ryu Hadouken at $072A80. Its first instruction is a PC-relative table fetch (move.w table(pc,d0.w),d0 / jmp table(pc,d0.w)), so the head stays vanilla — a six-byte JMP there clobbers the fetch. The isolated jmp $0135E6 at $072B14 already trampolines to CLEAN $39D0A8: parent $54, steal $37A, write $52, then the stock D1 walk. The spawn veneer at $39CED8 already copies $37A → $52 before it writes type $48. XvSF oracle vs Ryu is life 144→132, debit 12, special $08. That debit, on this zip, is the measurement I am still taking. Same veneer discipline. One more field.

The kiss holdsGrab veneer $39D0CA. Face, snap $24, JSR $39BCE0, JSR $0187EE, restore $a8. Public $06 vs Ryu $12.CONFIRMED IN GAME

group	$20 Power Drain
actions	$32 / $34
spacing	p1x=79 p2x=115 · $24
steal	$0012 Ryu
stolen QCF	special $08 at age 204

Public Dual Rogue colors in this harness. Shadow Rogue’s navy remap is the secret $26 match. The 100% pass continues on the same slice: stolen type $48 D1, life debit, and a Hadouken that survives spawn the way XvSF does.

MAME case=absorb snapshot 0002.png · live/absorb.log age 20 / age 204

The 100% pass

The 100% pass is one seating, taken further, until Dual Rogue is the fighter I play in XvSF — on a vanilla Euro parent, without expanding the ROM, and without copying her 68K onto $02F6FC.

Every closed slice on this chair used one method. Name the XvSF oracle. Find the MvC object that already speaks that language. Write a veneer that is small enough to live in CLEAN. Live-gate it. Keep the stock path that versus construction still needs. The kiss is that method: $0187EE is MvC’s attach, $39D0CA is the 94 bytes that face, snap, and restore. The stolen QCF is that method: $026780 came with the detector. Type $48 is that method in progress: vanilla head at $072A80, isolated trampoline at $072B14, steal already on $52 before the type byte.

After debit, the named slices are the same job. The following is the live remainder recorded on the historical early candidate, not a claim about exact V18b and not a short-harness promotion:

She still plays like Hulk. The button-to-group map on this chair is still Hulk’s, so Gamma Charge / Gamma Crush fire in her slots and her own specials do not. After a special she can glitch into playing like Ryu — stolen-id leftover from $026780, not a second character on the roster. The Street Fighter supers she should be able to use after Power Drain do not work. Her own supers do not behave. Matches still crash.

Getting her to XvSF level is the rest of the 100% pass: measure those oracles in X-Men vs. Street Fighter, write veneers that speak MvC, live-gate them, keep Hulk’s window $02F6FC parked. Copying XvSF 68K onto that window is not the path.

Assists and DHC at $07A242 / $09DA20 still carry Hulk’s partner kit because Rogue never had a row there. PCM $3A0..$3BF remaps the IDs, but audible exact- V18b voice/rate parity remains unaccepted. The complete source command census contains no persistent flight mode; jumps, super jump, air dash, dash links, and air attacks are the real movement requirements. An early short MAME harness PASS is historical evidence for its own hash, not an exact-V18b playtest.

What we are doing better now

The Jin draft treated expansion as the authorized next step after a red ledger. The better move was to spend a chair the Magneto plan had already priced and then not used.

The editor side is the same instrument as Magneto and Storm. Dest dump profile rogue-mvc keeps overlay 0, accept_label HULK, live id $06. Join v5 closed the four ETC cells that stock-shape mapping parked at MvC 29. Scoreboard: 196 HIT, 197 dest, names 100%, slots 100%. Details live in the harness post.

The 68K work is the same discipline taken further. An earlier pal-upload skip at $00A4CE greyscaled the body and crashed before versus, so versus upload stays the stock path: do not poke $FF3052. Hulk’s window $02F6FC stays parked. Her blob sits in CLEAN $399816 so opcode and data agree. Power Drain can kiss. SPECIAL 00 was the first live special on this chair, and it is still Hulk’s recognizer. The 100% pass is that same seating, continued until she plays like the Rogue in X-Men vs. Street Fighter: her specials, her supers, stolen Street Fighter supers after Power Drain, a match that does not crash or turn into Ryu. Assists and DHC, PCM $3A0, and source-exact jumps, super jump, air dash, dash links, and air attacks. The source census contains no persistent flight state. I keep taking her toward the fighter I play in XvSF. That is the project.

Magneto’s two-tree story and this one share a lab: the unified Magneto article, the nested HIT harness. The helper I always picked is on the roster cell. The square is her smiling XvSF bust, not a tiny idle body. Her 68K lives in CLEAN $399816. Exact V18b has real bounded passes, but its ledger still carries 79 OPEN/HOLD/FAIL blockers and no completion percentage. V28 establishes the 8/52 neutral read/prefetch baseline. Matrix M1 repeats it and adds only $CB50 during two exact active-CATCH configurations, moving the strict subset to 9/52. Those taps do not prove instruction retirement, one-to-one retry execution, caller closure, or behavior. The other 43 caller obligations and the complete state, exception, interrupt, retry, and return-epoch domains remain UNKNOWN. The 12-capture Matrix M2 tag experiment qualifies no additional caller: all eight tag variants remain HOLD because their lifecycle crosses a both-active interval and never reaches the required contiguous 60-frame settlement. It awards zero gameplay or release credit, and the Rogue download remains temporarily withheld. The locked 572-case native-move matrix, CATCH choreography, full moves and combos, audio, presentation, team/stock regression, and one same-hash manual run remain downstream. That is the rest of the project, not a caption.