Forsaken: original guest state-buffer selection, 2026-11-12

The original guest deliberately selects the additive buffer at 0x469275 when its float at [0x795048] differs from zero. All 620 retained selected calls agree with the original compare, x87 condition bits, integer TEST result, caller arguments, global handle and actual Execute object. No CPU or runtime defect was proved, so there is no production patch. Static float writers and vertex-color consumers support a death/respawn fade effect; texture provenance, exact sampling and target history remain unproved.

This advances the producer boundary left by the accepted 1715 result (a51eacd69) and retains the paused-worker constraints (eec039872). The root accepted the 1715 artifact review on the messageboard at 17:34:43Z. The old 1645 and 1715 pixel witnesses remain historical records, not numeric address or causal-input templates for this fresh run.

Original instructions and observed caller

Run: scratch/runs/20261112-forsaken-guest-state-selection-1735/ (ignored). Source base: 700fbb1ec61fbbea6f27fdcbed74b81bd2cd4522 plus the exact relevant binary patches in that run. Original media remained at test/binaries/candidates/forsaken-demo/installed/ForsakenDemo.exe, SHA-256 366ec5d0943daca839836296f9fb52724e94dc155630c6b0c30d7181106b5e24. Both actual CPU/GPU owners used normal module 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17.

Read-only original disassembly gives:

469262  fld   dword [0x795048]
469268  fcomp dword [0x4ee398]   ; observed bits 0, float zero
46926e  fnstsw ax
469270  test  ah,0x40           ; x87 C3
469273  jne   46928e            ; equality/unordered skips additive Execute
469275  push  1
469277  mov   eax,[ebp+0xc]
46927a  push  eax               ; viewport
46927b  mov   ecx,[0xa9d04c]
469281  push  ecx               ; selected Execute-buffer handle
469282  mov   edx,[ebp+8]
469285  push  edx               ; device
469286  mov   eax,[ebp+8]
469289  mov   ecx,[eax]
46928b  call  [ecx+0x20]         ; public Execute, return 46928e

The immediately preceding opaque Execute returns at 0x469262. After the selected call the guest calls 0x4d63f0; at 0x4692ab it compares the same float again and restores opaque state for equality/unordered (return 0x4692d6). Thus nonzero intentionally leaves additive state around this scene pass. Observed frame return 0x467147 led to original 0x467142 call 0x46909d; the caller supplies device and viewport. All 620 selected frames have that return, not an inferred historical caller.

The first retained selection is Execute 19907, boundary 357971, in diagnostic-tick-71.json.gz: float 8.236000061035156 (bits 1090766504), matching the sampled frame delta, versus zero. Actual EIP/previous EIP are 0x469275; previous-previous EIP is 0x469262. x87 SW 0x84 has C0/C2/C3 clear; TOP/tags are zero. Lazy integer flags are TEST (flag_op=3, flag_sign_shift=7, flag_res=0): JNE is not taken. Stale lazy flag operand slots are not interpreted as AH. EBP 122682468 contains caller 0x467147, device 136454216, viewport 136454224. Stack arguments are return 0x46928e, device, 136458472, viewport, 1, exactly matching the API. The actual code operand [0xa9d04c] contains 136458472: object slot541, entry131822496, buffer16734700, size48, instruction range0..40. Its raw header 08080400 contains (27,1),(19,2),(20,2),(21,2).

selection-analysis.json retains 7,520 relevant snapshots: 620 selected and 1,142 equal opaque-restoration calls, zero comparison/flags/branch/ argument/handle mismatches and zero observer errors. Equal restoration has float0, SW 0x4084, TEST result64. Later selected values reach float512. Flags are assessed at the selected/equal-restoration calls, where the argument pushes/moves preserve TEST, not at unrelated scene entries.

Why this state exists, and what remains unresolved

Original 0x4a46cb clears the float. 0x4a4cc3..0x4a4ccf and 0x4a4e75..0x4a4e81 add frame delta [0x794d20]; 0x4a4ce2 and 0x4a4e94 clamp to512. Nearby 0x4a4d6a decrements lives [0x54c144], with a respawn path resetting the float at 0x4a4daa. Vertex generation at 0x439e29 converts the float to intensity and folds values at256 using 512-minus-intensity; 0x43a531..0x43a558 adds/clamps/packs vertex color. 0x43a565 also branches on this float. This is static evidence of deliberate fade-related additive rendering, not a live trace proving every writer or the exact intended visual duration.

Two separate fresh witnesses are preserved completely in remaining-witness.json, including original Execute, state transitions, owned/public/replayed state, triangles, texture candidates and target:

Witness Actual source / producer Texture and target Pixel
First retained sampled white after combat-ready fence4825 GPU24138; source 7608/1; Execute19824, return 0x40502a; preceding state Execute19822 return 0x4693e9, handle136458464 texture244, entry131812992, DIB476602368, 256×128; target131805408, DIB471527424, 640×480 16bpp format1; ZWRITE0, ONE/ONE [41,166,189,255] → white
Scene after the requested selector GPU24337; source 7680/0; Execute20012 return 0x44182f; selected Execute20011 return 0x46928e, handle136458472, buffer16734700 texture210, entry131811904, DIB473559040, 256×256; target131805408, DIB472166400; ZWRITE1, ONE/ONE [255,223,231,255] → white

The first witness is in capped tick70 and precedes the first retained 0x469275 selection in tick71. Its event envelope says legacy-fence7609; the correct draw join is the record's 7608/1. The preceding sampled scene Execute19818 has float0. Do not attribute that first witness to this selector or to primary fire.

For the requested scene, Execute20009 completes opaque restoration, Execute20011 selects additive with float97.76699829101562, correct SW/TEST/caller/handle, then Execute20012 observes the same float. Triangle8 has interpolated UV (0.22651625968240824,0.15406722707995493) and vertex RGBA (183,245,221,183), (145,100,78,145), (183,100,78,183). Native level-zero candidates at (57,38),(58,38),(57,39),(58,39) decode to (140,97,82), (132,81,66), (165,105,90), (132,97,74) respectively. These are candidates, not an exact GPU sampling oracle.

The concrete remaining boundary is original texture asset/writer provenance into those actual DIB bytes, their uploaded/sampled texels, and ordered prior target contents for these particular source commands. The first witness's triangle13 samples a different texture near UV (0.3645227393621847,0.20933279251005968) with vertex RGB64; its candidates include white and (255,255,214). Texture/target/order correctness is not established by a white pixel or by matching additive state. No speculative blend override is justified.

Observer qualification and natural original run

The new run-local selection reader executes only at the existing Execute API observer. It validates actual original opcodes and derives their data operands, reads the actual owner's translator/memory/register exports and stack, and performs no guest call/write or getter that materializes flags. Inherited final1715 attach, pinned translator, Execute/return observer, identity, ACK cycle, readiness and independent release helpers are reused.

Before allocation, the emitted reader passed five bounded original-code cases (0,1.25,512,-1,NaN): observed/control whole 512MiB SHA, registers and block/control events match. The adapter preserves arguments, receiver, return and thrown-error identity even when observation fails. The actual Chrome production CPU/GPU bootstrap controls and updated CPU Encoder fixture passed; Encoder bytes/messages match the unobserved baseline. Unchanged negative controls are enumerated with hashes in reused-controls.json. Local provision/import and stale-pin negative controls passed before allocation; helpers/pins were frozen and hashed. All 1,559 remote provision paths and the final actual import gate passed before original execution. API provision batches remain separately capped.

The private original ran normal menus, Normal VOLCANO, four native Enter pulses, a600/700/300ms, then Control1800ms: eight prearmed independently released pulses, sixteen trusted edges, eleven full PNGs and eleven ACKs including ordinary quit. All captures match all 19 owner/module/epoch fields. Chrome's actual CPU target is 909CB5AA3E9445591331498E91FD0FE6; GPU target 11555E9168C4EBA6EF638454565D5236, endpoint1 WebGL2 Intel UHD620 ANGLE/Mesa. The corridor frame shows lives4/respawn before primary fire; primary combat still renders a room. Both later passive captures show a white viewport, first complete at 17:45:00.415Z. This differs from historical selected pixels and establishes no causal input attribution.

The retained 153 drains contain 607,813 boundary rows, 427,973 CPU rows and 9,233 GPU events. All 39,932 state applications agree;34,306 Execute entries have 34,306 returns. Limits: 12,696 startup CPU drops, 29 ground record caps (including first retained white tick70), three unjoined tail replay keys at 12448/0,12449/0,12449/1. Boundary rows have no gaps/drops/errors; GPU events have no reported drops. These limits preclude a global-first-white claim.

Recovery and shutdown

A fresh observable independent two-hour watchdog was armed before creation. bx_cjkc84d4 was created17:37:44Z with initial expiry19:37:43.027Z. Actual free /home was sampled above1.5GB during input; there was no RAM gate. Shared growth consumed the32MiB reserve; no more gameplay input followed the 17:45:19Z near-floor report. Later shared-tree floor crossings are retained on the messageboard, not represented as a continuous storage guarantee.

All 297 unique remote artifacts passed two independent GETs plus local SHA verification before stop. Source recovery is exact Git SHA plus binary patches, media path/hash; no source/media archive was created. Closed large logs/patches are losslessly gzip-packed with both hashes; duplicate PNGs are hardlinked only after independent retrieval and verification. validate-closeout.py rechecks retrieved SHA, frozen helpers, lossless recovery, eleven complete PNG CRC/inflate/hash checks,19-field identities, input/ACKs, selection results, actual parents and normal archive.

Ordinary quit completed17:46:12Z. Actual Chrome parent and browser parent returned0 with separate stdout/stderr EOF; outer8048 is known exited0 with untruncated output. Local lifecycle client/owner/watchdog actual waits are0 with captured EOF where piped. Separately, the unified exec session68586 reported143; lifecycle-exec-return.json preserves this unexplained wrapper status. Its cause is UNKNOWN; no all-local-parents-zero claim is made.

Final box state ARCHIVED, expiry null. Stop stop_22369c81dbcc8442 completed 17:47:15.111Z, forcedfalse, errornull, after verified recovery. No box was resumed/deleted. No production code changed, hence no production regression/build or matched patched-original claim. The useful next work is the specific texture/target/order witness above, retaining normal timing, actual-owner address discovery and the bounded read-only observer controls.