Constructor (1997) original DOS demo

Requested in Telegram message 63 from https://x.com/oldyzach/status/2110085411389469074. The acquired release is the original 20-minute demo, not the retail game or a remaster. Although the Archive item labels it Windows, this package's executable is DOS. No native Windows demo has been verified.

Original media and executable

The payload is an ignored local fixture. On bx_yqnspudj, acquisition and source-only scan artifacts are under /home/user/constructor-preflight-20261013/; the extracted tree is under corpus/test/binaries/dos-games/constructor-demo. Local metadata does not imply the payload has been copied onto the dashboard host. At the initial preflight the coordinator's test/binaries/dos-games/constructor-demo/GAME.EXE was absent; see the 2026-11-01 recovery below for current local state. The dashboard now checks the local entry of blocked titles too, reporting this missing path separately from the retained compatibility blockers. That presence check is not a full payload hash verification.

Intended original route, not observed gameplay

README.TXT explicitly identifies a twenty-minute timed demo and instructs running GAME from its installation directory (C:\CN). Press Enter through the splash screens, choose PLAY GAME, select a team color and enter a player name, then ACCEPT. The documented scene begins at HQ with foremen. Its first construction task is laying the lumber yard foundation, assigning a work gang, then staffing the completed yard to produce wood. These are instructions from the original media; none of those screens or actions has been observed here.

Initial ToyVM dependencies

The source assessment uses main 2d4ca864e and the existing tools/toyvm-dos-corpus.js, without executing a guest:

  1. Directory-aware file access: five distinct original resources are named DATA.FIL, in DATA, FLICS, GFX, LFX, and SFX. Their sizes and hashes differ. ToyVM's tools/toyvm/dos.js hostPath() drops directories and looks up only the basename. Flattening this payload loses required assets. The duplicate root and MUSIC/DOS4GW.EXE is a second basename collision.
  2. Title-specific extender qualification: GAME.EXE uses external DOS/4GW. The generator retains its unqualified extender assessment. ToyVM has paging and has reached protected mode for other programs; this is not a claim that every protected-mode game is unsupported or that DPMI caused a failure. Constructor has no native attempt or observed startup failure yet.
  3. Browser mouse delivery: the current DOS browser route wires keyboard input; its canvas pointer handler focuses the canvas rather than forwarding guest mouse coordinates/buttons. CLI INT 33h support alone does not supply the original README's mouse-driven menus and construction controls.

No CD-ROM or VBE 2.0 requirement was inferred merely from the game's age or platform. No silent service stubs, guest changes, or renamed resource workaround were used. The shared DOS dashboard row links provenance and these notes and remains blocked with no launch route or gameplay claim.

Validation

The original archive was extracted and scanned on bx_yqnspudj. The selected generator and --check --only=constructor-demo agree. All 28 file sizes/hashes are recorded, including the LE program and distinct resource files. The existing 22 corpus scanner/assessment checks pass. Dashboard view-model validation checks the Constructor row, provenance, file list, notes, missing local payload, and absence of a launch URL. All 14 dashboard corpus/view tests pass, including a regression that adds a missing blocked entry and confirms compatibility remains blocked even once the local entry exists. Initial sparse-checkout dependency failures were resolved by including the original dashboard modules and rerunning. Source-only receipts are retained in scratch/runs/20261013T2354Z-constructor-corpus-preflight/. No emulator, build, or browser was launched; screenshots, audio, performance, gameplay, and saving remain untested.

Directory filesystem source work (2026-10-14)

The initial basename blocker was reproduced with the real Machine file API on db2e8905a: five synthetic directories containing different DATA.FIL bytes all read the root DATA.FIL instead. Without a root copy they miss. This is a filesystem defect, independent of the game's DOS/4GW startup.

The source change mounts fileRoot at C:\ and keeps complete relative names through openFile, whole-file/EXEC reads, create, copy-on-write, temporary files, delete, rename, attributes and FindFirst/FindNext. Case-insensitive component lookup retains the existing trailing-dot/CP437 blank compatibility. Slash and backslash separators, . and .., root-relative and C:-relative paths resolve against an actual current directory. .. at root stays at root; host symlinks are outside the mounted namespace. AH=3Bh changes to an existing directory; AH=47h returns it without a drive or leading slash. C: remains the only mounted drive; another drive's paths fail and its AH=47h request returns error 15. AH=0Eh reports three as the highest drive index plus one and leaves C: selected for unavailable requests. Directory creation/removal and other unimplemented services remain unimplemented.

Guest overlay keys now contain their directories. An open original-file handle remembers its resolved path if cwd changes before a write. Writes do not mutate host originals; deleting or renaming originals (including shadowed originals) returns access denied instead of falsely succeeding or revealing old bytes. Missing directories no longer alias root. Two old rename fixtures were adjusted to create/use the directory they had previously expected to be discarded. FindFirst stores a directory token in the DTA reserved bytes, preserving copied and nested searches across cwd changes; repeated searches reuse directory tokens. CLI --save-files preserves relative overlay paths and refuses traversal and symlink destinations. Root-only callers retain root-only map keys.

The same fix includes both generated browser bundles and their filesystem shim, the dashboard manifest validator and the browser loader. Ordinary files retain their complete manifest paths. Ambiguous case/dot aliases, path traversal and file/directory conflicts are refused. Previously declared requiredOriginalFiles overlays retain their explicit guestPath (for example the original U4 save templates); preserving these mappings does not restore implicit basename aliases. Entry programs retain their manifest path too.

Validation is pure JavaScript: 24 filesystem/API groups across Node and both shipped bundles, 15 existing temporary-file groups, 14 existing rename groups, 19 dashboard loader/server tests, and the overlay export contract pass. The bundle reproducibility check and git diff --check pass. The required-original file test's synthetic portion passes; its original U4 fixture portion is explicitly skipped locally because that payload is absent. The initial sparse bundle generation lacked fonts/Terminal.fon; adding the tracked font to the isolated checkout resolved that setup failure before generation succeeded.

These checks execute no guest and establish no Constructor startup, input, audio or gameplay result. The original 28-file fixture was not modified or re-extracted. The catalog remains blocked pending actual original DOS/4GW startup and browser mouse delivery, followed by the ordinary menu-to-HQ and building-controls route. No demo limit was changed. A later serial runtime validation should run test/test-toyvm-dos-rename.js --native, test/test-toyvm-dos-files.js and test/test-toyvm-browser-bundle.js, then use the pinned original Constructor fixture; those guest/native checks have not run for this change.

Environment-MCB startup blocker (2026-10-14)

The first original GAME.EXE attempt on full-path checkpoint 9a74aad7d spent 200M dispatches in a blank mode-3 screen. Its two dominant INT 21h return sites were 110:1089 and 110:22fa. Static MZ disassembly (96-byte header) maps them to AH=49h free and the AH=52h list-of-lists wrapper: the loader walks MCBs, frees blocks owned by its PSP except retained blocks, then restarts the walk.

A bounded observer confirmed 5,714 AH=49h calls for ES=0060 paired with 5,714 AH=52h calls in just 1M dispatches. The environment MCB owner stayed 0100 after every purported successful free. installArena wrote the initial environment's header, but never enrolled it in memBlocks; AH=49h therefore did nothing, and mcbSync reconstructed only the arena above the PSP.

The fix enrolls the initial environment as an ordinary allocation and reconstructs the arena from that block while preserving the intervening DOS-owned block. It permits real freeing, splitting, reuse and coalescing. The focused old-code regression fails on the unchanged owner; the candidate's seven MCB groups pass. EXEC environment/boundary/resident tests, DOS files/rename/termination tests, and all 24 pure-JS full-path tests pass. Two existing EXEC failure tests now compare allocator snapshots rather than assuming the initial environment is absent from the allocation map. The same updated EXEC fixture also passes against the parent. Both browser source bundles were regenerated and pass freshness checking.

The original candidate now frees segment 0060 once, changing owner 0100 to zero, and makes only three list-of-lists calls. It then exits with the visible message Not enough memory on exec, before any AH=4Bh call. It has not reached DOS/4GW protected mode, a menu, HQ, or building gameplay. This is a resolved startup loop and a newly exposed loader-memory blocker, not gameplay support. The next bounded diagnostic should record AH=48h/4Ah requested and returned sizes, carry flags, MCB boundaries and the loader's required size; do not increase memory or bypass the guest check.

Both runs hash-checked all 28 originals before and after. No guest changes, memory/RNG patches, input injection, browser run, or silent service stub was used. The old arm had a 1M-dispatch bound and the candidate a 30M bound; the latter exited early. Captures were reviewed: blank parent and the candidate's memory-error text. Evidence, including the initial fixture failure, corrected tests, process cleanup and bounded observer, is in scratch/runs/20261014T0055Z-constructor-environment-mcb/. The tested source depends on the still-separate full-path change 9a74aad7d. The initial investigation's unmet mouse/gameplay requirements still apply.

PSP contiguous resize correction (2026-10-14)

The next bounded original comparison confirmed the memory-error cause. After releasing the block at 0443, the loader requested AH=4Ah on PSP 0100 with BX=FFFF. The parent reported 40448 available paragraphs despite a retained four-paragraph block at 9EFC. Retrying that size returned success but left the PSP MCB at 834 paragraphs: the PSP-specific resize path moved the global frontier without consuming the adjacent free region.

The correction derives the current PSP size from its MCB, consumes only adjacent free space when growing, returns a shrunk tail to the allocator, and stops at the next held MCB. An oversized request grows to the contiguous maximum and reports CF/AX=8/BX=that maximum, consistent with FreeDOS DosMemChange. No memory ceiling or guest byte changed.

Three new regressions prove grow/shrink/reuse, the maximum before a held neighbor, and refusal to borrow disconnected free space. The exact previous 23670bef80 DOS module fails the actual-MCB-size assertion. All ten MCB groups, eleven EXEC environment groups, EXEC boundary/resident, DOS files, rename/termination, 24 path checks, native rename across four dispatch variants, browser-bundle VM tests and bundle freshness pass.

The candidate original request now reports and actually owns 40443 paragraphs, preserving the block at 9EFC. It reaches INT 21h AX=4B03 (load overlay), then exits with Bad format on exec. The DOS handler currently rejects EXEC subfunctions other than AL=00/01 with AX=1; AL=03 is the next concrete missing behavior. No protected-mode, menu or gameplay claim follows from this progress. Both text captures were reviewed; all 28 original hashes match before and after each run. All 34 observed PID and process-group checks were independently absent at release.

Evidence: scratch/runs/20261014T0106Z-constructor-psp-resize/. The matched parent DOS module was loaded from immutable commit 23670bef80 through a diagnostic Node require preloader; other runtime source files were unchanged.

EXEC overlay loading and cache replacement (2026-10-14)

INT 21h AX=4B03 now loads an overlay at the caller's segment with the independent relocation bias from ES:BX. It preserves the PSP, environment, arena and caller control flow. MZ/ZM loading validates the declared body and relocation table before any writes, excludes appended data, and refuses out-of-range relocation targets or destinations. Flat overlays load at offset zero with the DOS 0xFFFE transfer bound, matching FreeDOS task.c. Successful loading flushes translated code and releases live-uop heads so reloading the same address executes the replacement bytes.

The eight API groups fail against immutable parent 56d339a15 and pass on the candidate. Native reload tests return distinct 1234h/5678h values across all four dispatch variants. Disabling invalidation produces the expected stale-code failure. The live-uop variant installs a hot arithmetic head and releases one head on replacement, then returns the second value. Its initial short fixture installed no heads and failed the coverage assertion; that failure remains in the evidence, and the corrected hot-loop fixture passes. The filename-based unit tier includes the new test: tools/test-tiers.js --check reports 1892 files, 1484 unit, 402 e2e and 6 smoke. MCB, EXEC, DOS file/rename/termination, 24 path, native rename, browser VM and generated-bundle checks also pass.

Original GAME.EXE now loads DOS4GW.EXE rather than printing Bad format on exec. Its 30M-dispatch bounded run remains in real mode with a blank mode-3 display; it has not reached a menu or gameplay. An enriched observer finds CS:IP 02CD:6309, CR0=10h, with no trap, unsupported opcode or protected-mode stop. The initial observer's vm.get('ip') was an interpreter field; only the enriched vm.get('gip') value is the guest IP. Both artifacts are retained.

Original disassembly identifies a search for duplicate IVT far pointers below vector BFh. DS is genuinely zero (the instruction is not relocated). The scan compares each earlier vector against later vectors and restarts after SI wraps. Current initialization assigns every vector a distinct F000:(0100+vector) target, unlike BIOS shared default IRET handlers. This is the next source-backed conformance lead; no IVT change was included in this overlay correction. AH=38h was unhandled once but is not established as the scan's cause.

Evidence: scratch/runs/20261014T0123Z-constructor-exec-overlay/. Original 28-file hashes remain unchanged. Parent error and candidate blank captures were reviewed. The initial suite, enriched observer, failed coverage attempt and passing hot-uop attempt have 60 independently absent PID/group checks. No guest bytes, memory ceiling, RNG, browser or gameplay bypass changed.

Shared default interrupt handler (2026-10-14)

The matched original observer confirms DS=0 and the duplicate-vector scan at 02CD:6309. The parent's IVT has distinct far pointers; its 1M budget expires inside that search. BIOS initialization normally gives otherwise-unused vectors a shared IRET handler, then installs service/IRQ handlers; see SeaBIOS ivt_init. The emulator now supplies an actual shared IRET instruction for unused software vectors, preserving all 18 implemented service vectors, hardware entries and defined-but-unimplemented BIOS/DOS services on their diagnostic host paths. The dispatch loop recognizes only the actual host-entry range. Hook detection compares the complete far pointer, including guest hooks in the ROM segment.

Four API groups fail on immutable overlay parent 567dc3bb3 and pass on the candidate. Four native variants prove flags, stack and BX preservation, DOS get/set-vector save/restore, and guest-hook chaining through the saved IRET. An actual INT13/AHFF also must reach both existing unhandled diagnostic counters; the unused vector must not reach host service dispatch. Word-register and protected-mode timer regressions, overlay reload including hot-uop release, MCB/EXEC/file tests, 24 path groups, browser VM, bundles and tier checks pass.

The first native fixture failed: startup PUSHF was F203h and actual IRET returned 7203h, differing only in reserved bit 15. Two diagnostic attempts preserve that finding. The corrected synthetic program explicitly POPFs valid 0202h before setting carry and comparing full flags. The same corrected fixture is used for the parent/candidate comparison; no emulator flag behavior changed.

Original GAME.EXE now passes the scan and exits after 306734 dispatches with DOS/16M error: [26] 8042 timeout. CR0 remains 10h (real mode), with no trap, unsupported instruction or protected-transfer stop. The next source lead is the keyboard controller/A20 protocol, not a reason to patch the guest or skip the extender's hardware check. AH38, INT15/BF, INT2F/16 and INT15/C0 diagnostics are retained and are not yet established as this timeout's cause.

Both original runs verify all 28 hashes before and after. Blank parent and candidate timeout captures were visually reviewed. All 66 observed PID/group checks are independently absent. Evidence, including the initial failed fixture and corrected suite, is in scratch/runs/20261014T0133Z-constructor-default-ivt/. No protected-mode startup, menu, HQ or building gameplay has been established.

8042 output port and A20 gating (2026-10-14)

Original DOS4GW file offset 29A1 waits for controller input-buffer-empty, writes D1h to port64h, waits, writes the output byte to port60h, then waits. The helper at29F8 tests status bit1 up to65536 times before reporting error26. The emulator had no port64 handler, so its default FFh made IBF permanently full; the matched parent records exactly65536 status reads.

The controller now consumes D0/D1 synchronously, reports IBF/OBF from its state, and maintains an output latch. D0 queues a real reply; D1 consumes one subsequent data byte. Keyboard bytes and the IRQ1 data latch survive controller replies. A20 changes the actual address mask: off clears bit20 alone while preserving higher address lines, on uses the existing16MB address space. Existing XMS allocation/lock openBus calls use that same enable path, keeping controller readback coherent. No memory size changed. Existing XMS03–06 nominal enable/disable/refcount behavior remains outside this change.

The first original attempt failed explicitly on command FFh. This is the documented pulse-none command, so its no-op output behavior was added and tested; unknown commands and CPU reset requests still fail explicitly. The command and status definitions are supported by QEMU pckbd.c. The initial failure and exact preceding source/test remain in the evidence.

Seven API groups fail on immutable parent68ac64f2b and pass on the candidate. Native tests across four dispatch variants prove FFFF:0510 aliases0000:0500 when A20 is off and becomes distinct when on, while00300500 aliases00200500 rather than00000500. They also preserve the stored bytes across toggles. Controller sequencing/status/readback, keyboard preservation, explicit failures, and existing XMS allocation/lock interactions pass. IVT native/diagnostic checks, word-register and PM timer regressions, overlay hot-uop replacement, allocator, EXEC, DOS files,24 path checks, browser VM, bundles and tier checks pass.

Original GAME.EXE performs D1/DF then FF, leaves the output latch DFh with A20 enabled, and enters protected mode (CR0=11h). It stops after633142 dispatches at the existing stuck-code guard, CS:IP=3C78:FFFFA7E2, without traps, unsupported instructions or a protected-transfer stop. The mode3 screen remains blank. The next issue is the far control-transfer/address state at that stop; there is no menu, HQ or building-gameplay claim.

Both matched original arms preserve all28 hashes. The parent log records the full8042 timeout; its reviewed console PNG visibly shows only the DOS/ prefix, so the log is the complete text evidence. The candidate reviewed PNG is blank. All92 observed PID/group checks are independently absent. Evidence: scratch/runs/20261014T0141Z-constructor-8042-a20/. The corrected suite contains 21 passing commands including the expected parent API failure.

Protected near-branch target width (2026-10-14)

A passive observer on controller checkpoint50544aa9e captured a valid protected D=0 CS descriptor: selector3C78, base1348E0h, limitB2FFh. Execution went from offset95EAh to FFFFA7E2h in one handback. The decoder then refused bytes FEFF at the incorrectly wrapped linear address;202 repeated no-work handbacks triggered the stuck guard. This was not a timed guest loop. The observer did not resolve the individual branch inside that3630-dispatch handback.

decodeOne used its fetch-width helper for relative near-branch destinations. Protected D=0 code needs wide fetch offsets when its descriptor allows them, but a16-bit near transfer truncates the target independently of that fetch width. The correction separates target width for JMP rel8/16/32, ordinary Jcc rel8/16/32 and CALL rel16/32, including66 overrides. CALL return fixups match the pushed operand width. Fetch and untaken conditional fallthrough remain wide; LOOP and JCXZ are unchanged. See Intel volume2A JMP/Jcc/CALL.

Four decoder groups prove16-bit underflow/overflow,66 overrides in both D-bit defaults, high-EIP D=0 fetch and wide untaken fallthrough, and the 16-bit CALL return fixup. The unchanged test fails on parent50544aa9e. Four native dispatch variants execute protected16-bit backward/forward wrapping jumps and a wrapping CALL/RET, verifying the reached code and result. The existing8042 native aliases, IVT hook/diagnostic checks, overlay hot-uop reload, PM timer/word-register, allocator, EXEC, DOS file/path, browser VM, bundle and tier checks all pass (22 commands including expected parent failure).

Original GAME.EXE now prints the reviewed DOS/4GW Protected Mode Run-time Version1.97 banner. It remains in mode3 with CR0=11h and spends the30M dispatch budget without a stuck, trap, unsupported-opcode or protected-transfer stop. The hottest observed entries are80:323E,80:06FC and80:0954, about31K each; the final address is80:0C34. This budget stop is not evidence of a freeze, and no menu, HQ or building gameplay has been established.

All28 original hashes match before/after both arms. Parent blank and candidate banner PNGs were visually reviewed. All48 suite PID/group checks and the six earlier passive-observer checks are independently absent. Evidence: scratch/runs/20261014T0151Z-constructor-near-width/; the prior passive capture is included as its pm-stop subdirectory. No guest state, RNG, memory size or stuck-guard behavior was patched.

2026-10-14: EXEC preserves held DOS allocations (UVCONFIG)

The unchanged 200M-dispatch / 30-second run on 11e8935 stopped at 41,597,432 dispatches after ordinary initialization and SFX loading, with protected RETF type rejection at CS80:0C78 targeting selector005C. This was not a budget freeze. Three captured code signatures independently map CS80 offsets to DOS4GW.EXE file offset206324 + IP. The transfer sequence at0C74 is operand-size-32 PUSH EDX, PUSH EAX, RETF; earlier instructions read the callback offset/selector from ES:[DI+2]/[DI+6].

A separate bounded passive capture found valid shared LDT/GDT base110010h, LDT limit3FFFh/access82h. Descriptor005C bytes6700741E00810000 describe a 286 TSS, so RETF correctly refuses it. The callback table at ES=A8, base7E60h had instead been overwritten by the environment strings for C:.\UVCONFIG: the apparent offset is ASCII H=C: and selector005C is the backslash. EXEC chose environment07E6 because allocTop equaled the DOS ceiling, ignoring an existing128-paragraph allocation at07E6. It replaced that tracked block with a4-paragraph environment and put PSP07EB inside it.

EXEC now searches a whole environment/MCB/PSP/minimum-image gap that excludes every held block and its header. Child max allocation stops at the next held MCB. Existing free holes are split around the reserved interval, and the global high-water mark stays above live allocations. The legacy imageTop fallback for never-shrunk launcher stubs remains; this is not a wholesale redesign of EXEC allocation strategy. RETF validation is unchanged.

Four added groups in the already-tiered exec-boundaries test cover the default-ceiling held block, fragmented gaps and max-allocation cap, atomic no-fit failure, load-only placement, and subsequent allocation refusing to hand out child/held bytes. Parent11e8935 fails the held-block assertion; candidate passes all8 boundary groups and the21-command suite of existing allocator/EXEC/path/interrupt/8042/IVT/overlay/near-width native regressions, bundle regeneration/reproducibility and tier membership.

The original now reaches43,307,267 dispatches and exits without a protected transfer stop, trap, stuck guard, or unsupported operation. The reviewed console capture reports "No Graphics hardware detected by UniVBE! (0x0)"; the framebuffer is blank. This is a graphics-detection failure, not verified menu/HQ/building gameplay. All28 original pins remain intact. Independent cleanup found6 PID/groups absent for each passive run and46 for the suite. Evidence: scratch/runs/20261014T0205Z-constructor-exec-held (when disk guard permits export); remote /home/user/constructor-exec-held-evidence-20261014 contains the prior startup-progress and RETF-stop captures, scripts, pins, bounded logs, images and cleanup receipts. Port counts are limited to60/64/92; the first passive observer did not capture LDT fields, corrected in the second by using generated mget_ldtb/mget_pm_ldt_* exports.

2026-10-14: graphics contract and optional Trident experiment

Source3415419, unchanged original GAME launch: the passive default-VGA run repeats43,307,267dispatches, UVCONFIG exit6 then parent exit1. Parent issues three successful VBE4F00 queries and mode-info queries100/101; child issues two more successful4F00 queries. All controller replies are0102. There are no VBE mode-set or protected-interface requests before failure. Capped port census shows chipset probes and PCI configuration scanning; not a VBE query failure. The host returns unsupported chipset BIOS queries unchanged.

Static GAME.EXE file offsets (not runtime addresses) establish why simply implementing VBE2 would not avoid this path. At034E2A bytes6A016A01 push two true arguments, then034E34 calls0A2BCC with the path argument. That function calls version-query0A28D8 at0A2C27. At0A2C57 it tests its second stack argument [ESP+164h]; only if zero does it compare the returned version with0200h at 0A2C61 and return status2 for an existing VBE2 implementation. The sole direct caller sets that ignore-existing-VBE argument to1. Later034F8A checks a minimum runtime version0102h. Exact bytes/disassembly are in game-graphics-branches.txt. Independently, the primary VESA2 specification requires4F0A's32-bit protected interface: changing the advertised version alone would be an inaccurate capability claim. Reference: https://www.phatcode.net/res/221/files/vbe20.pdf (section4.13).

Two bounded original experiments select the existing documented svga=trident option; emulator source and all28 original files remain unchanged, no input or guest-state edits. Its existing API contract test demonstrates CRTC1F readback and independently persisted64KiB banks through sequencer0E, rather than just changing a vendor string. Both original runs reach156,076,089dispatches: UVCONFIG returns0, parent reaches sound configuration then exits1 with Problem reading SETSND.CFG. BUT generated UVCONFIG.DAT and UNIVBE.DRV identify InteGraphics SuperVGA / IGA168x internal DAC and clock, not Trident. This is evidence of incomplete hardware probing, not validated graphics compatibility, and no graphics mode or gameplay was observed. Retained generated files are host evidence only; they were not installed into the original fixture or reused by a later run.

Current Trident implementation supplies a ROM identification string, CRTC readback and one bank selector. It has no old/new extended-sequencer register banks, chip-version read side effects, or complete protected/extended register and mode-programming model. Generic sequencer and CRTC indexes are masked to0F/1F with writable array backing. UVCONFIG genuinely probes extended sequencer A7/C5, which alias ordinary indexes7/5 here. That alias is an observed model limitation; it is not yet proven the cause of the final InteGraphics selection. Generated driver identities therefore prevent accepting the option as a correct chipset solution. A bounded future hardware correction must test chip detection rejection as well as accepted chip identity, bank/window persistence, old/new register access, and actual mode geometry/pixels; a full SVGA model is larger than an identity patch.

The original manifest has no SETSND.CFG. Its SETSOUND.BAT invokes MUSIC/SNDSETUP.EXE in MUSIC then returns to the parent directory. Once graphics hardware behavior is credible, the ordinary sound setup workflow is the next legitimate way to create sound configuration; no synthetic configuration or audio bypass was inserted.

Combined evidence scratch/runs/20261014T0217Z-constructor-graphics-contract: default passive capture, two Trident captures, generated driver hashes and bytes, static branches, existing-model conformance test and reviewed console images. All22 independently checked PID/groups are absent. No additional emulator change is proposed from these experiments alone.

2026-11-01 original fixture recovery and next executable dependency

Run: scratch/runs/20261101-constructor-original-fixture-recovery/. Static/acquisition assignment only; Urban owns runtime/source/build. The fresh public DESKTOP_APPS export in lib/apps.js has 46 entries and no Constructor entry (full list, source hash and HEAD retained in public-list.json). This checks the current public picker definition, not a remote deployed website.

The exact original URL above returned HTTP 200 after one Archive redirect in 5.193630 seconds, with no retries. Actual downloaded bytes are 32,062,553 and SHA-256 5d764f6d2f536d6aae6fd18122bc9bf2be20e84f101cbcdeafb1bcbd20eeed18. Every ZIP member was checked before persistence: no absolute/traversing paths, backslashes, drive prefixes, symlinks or special files; the 28 regular names exactly equal the checked-in manifest. ZIP CRC, size and SHA-256 checks pass for every file, totaling 63,205,318 bytes. Directory names/case are original; the five DATA.FIL resources and both DOS4GW.EXE paths remain separate. GAME.EXE matches d0ac2f05b9a813dc723f170f7418fbfd9af269abcbce8905062c3b63f27dc587; its MZ/LE identity and original sound-setup batch also match the earlier notes. No substituted executable, generated configuration or retail payload was used.

The complete fixture is now persistent at test/binaries/dos-games/constructor-demo/, with all 28 local sizes/hashes rechecked after writing. Transfer waited for Urban's exact author 2026-11-01T20:42:58+00:00 RELEASE runtime/source/build statement, terminal receipts and direct absence checks of its reported runtime PIDs. Free disk was about 2.18 GB before transfer and 2,113,626,112 bytes immediately afterward, above the 2 GB floor. No thresholds or other owners' artifacts were changed. Only the verified payload persists; the downloaded archive's own tmpfs copy was removed after local verification. The two bounded release-watch helpers were explicitly interrupted and reaped (exit130; actual waits84.975262 and 27.330045 seconds); the final watcher predicate/receipt accepts the explicit author RELEASE without waiting for its later note-publication DONE. An initial persist attempt failed before creating files because the dos-games parent was absent; creating the required parent corrected this acquisition-only issue.

Current source review (hashes in source-review.json) changes the interpretation of the early dependency list:

Once the fixture is persistent and a future runtime assignment is authorized, the existing CLI can mount its directory from the executable's parent:

timeout --signal=TERM --kill-after=5s 90s node tools/toyvm/run-dos.js \
  test/binaries/dos-games/constructor-demo/GAME.EXE \
  --dispatches=200000000 --seconds=30

This is a proposed bounded default-hardware startup diagnostic, unexecuted here. It needs the ordinary ToyVM runner dependencies and serialized runtime ownership; it does not need a replacement EXE or a new registry entry. The existing focused graphics regression is test/test-toyvm-vbe.js (guest/native, not run here): it checks served/refused VBE modes and rendered banked pixels, but does not qualify UniVBE chip identification or extended sequencer banks. A future hardware correction also needs focused detection/extended-register coverage and the unchanged original startup. Only after credible graphics is the original MUSIC/SNDSETUP.EXE workflow the next sound-configuration dependency; SETSND.CFG is absent from this authentic manifest. No setup was run.

node tools/toyvm-dos-corpus.js --check --only=constructor-demo passes: 28 files, 60.3 MiB, generated metadata up to date. Its extender,flat-fs-collision labels remain the stored static assessment; the collision label contradicts current directory-aware source and the passing real-payload host API check. These labels were not rewritten in this note-only assignment and are not new runtime results.

Recovery recipe from the repository root (only after the live runtime owner's exact release and with disk staying above 2,000,000,000 bytes). Check ownership first; preserve any existing destination. Keep the archive on tmpfs and persist only fully verified members; no compilation container is involved:

constructor_stage=$(mktemp -d /dev/shm/constructor-original.XXXXXX)
curl --fail --location --max-redirs 5 --connect-timeout 15 --max-time 180 \
  --max-filesize 33000000 --retry 0 \
  --output "$constructor_stage/CONSTRUC.zip" \
  https://archive.org/download/CONSTRUC/CONSTRUC.zip
python3 - "$constructor_stage/CONSTRUC.zip" <<'PY'
import hashlib, json, os, stat, sys, zipfile
from pathlib import Path, PurePosixPath
archive = Path(sys.argv[1])
assert archive.stat().st_size == 32062553
assert hashlib.sha256(archive.read_bytes()).hexdigest() == \
    '5d764f6d2f536d6aae6fd18122bc9bf2be20e84f101cbcdeafb1bcbd20eeed18'
pins = json.loads(Path('test/toyvm-dos-corpus/files/constructor-demo.json').read_text())['files']
expected = {f['path']: f for f in pins}
assert len(expected) == 28 and sum(f['size'] for f in pins) == 63205318
dest = Path('test/binaries/dos-games/constructor-demo')
assert not dest.exists() and not dest.is_symlink(), 'Preserve existing files'
assert all(not p.is_symlink() for p in dest.parents)
def free():
    s = os.statvfs('.')
    return s.f_bavail * s.f_frsize
with zipfile.ZipFile(archive) as z:
    for i in z.infolist():
        p = PurePosixPath(i.filename)
        assert not p.is_absolute() and '..' not in p.parts
        assert not any(c in i.filename for c in ['\\', ':', '\0'])
        assert stat.S_IFMT(i.external_attr >> 16) in (0, stat.S_IFREG, stat.S_IFDIR)
    entries = [i for i in z.infolist() if not i.is_dir()]
    assert len(entries) == 28 and {i.filename for i in entries} == set(expected)
    for i in entries:
        f = expected[i.filename]
        data = z.read(i)  # checks CRC and complete decompression
        assert len(data) == i.file_size == f['size']
        assert hashlib.sha256(data).hexdigest() == f['sha256']
    assert free() > 2000000000 + 63205318 + 16*1024*1024
    dest.mkdir(parents=True)
    for i in entries:
        assert free() > 2000000000 + i.file_size + 1024*1024
        out = dest / i.filename
        out.parent.mkdir(parents=True, exist_ok=True)
        with out.open('xb') as stream:
            stream.write(z.read(i)); stream.flush(); os.fsync(stream.fileno())
        assert hashlib.sha256(out.read_bytes()).hexdigest() == expected[i.filename]['sha256']
assert free() > 2000000000
print('Verified authentic 28-file closure, 63205318 bytes')
PY
# After successful verification only, remove this acquisition's own tmpfs copy.
rm -- "$constructor_stage/CONSTRUC.zip"
rmdir -- "$constructor_stage"

Run-local acquisition headers/timings, all 28 verification rows, static source pins, pure contract logs and actual helper waits are retained under the run ID. No guest run, build, browser, boat, deployment, EULA, gameplay, FPS, audio or qualification claim. Box ID/state/expiry: N/A; coordinator lifecycle untouched.

Restored original default startup (2026-11-01)

Run 20261101-constructor-original-startup executes the restored original once on local native Node v24.18.1, after Urban's exact 058aed059 author release, terminal launcher exit0 and fresh direct absence checks for all40 reported PIDs. The runtime/source/build claim was serialized; the other worker was restricted to synthetic inspector tests. No source, test, registry or build file was changed, and no project build was run.

Observed boundary: UVCONFIG returns6 with “No Graphics hardware detected by UniVBE! (0x0)”; GAME.EXE then exits1 at 0549:26EB. The CLI reports43.3M dispatches (rounded),113802 handbacks and4.32 guest seconds, ending in mode3. Its OS process exits0 because the runner completed normally; that is distinct from the guest's failure. startup.png was emitted and visually reviewed: the 80x25 text console contains the graphics error followed by initialization/SFX messages. Text overlaps on the first two lines, as recorded in both PNG and console dump. There is no usable menu, intro or gameplay screen, and no inputs were supplied. No FPS or audio result follows from the runner's counters.

Identity was pinned before guest execution and rechecked afterward:

The actual command in command.json retains the documented default hardware, timeout --signal=TERM --kill-after=5s 90s, --dispatches=200000000 and --seconds=30, adding only the identity observer and existing diagnostics: --trace-int --trace-fault --trace-entry=40 --trace-entry-regs --trace-io=3c4,3c5 --report --text --png=.../startup.png. No alternate EXE, --svga, VBE-version substitution or guest patch was used.

The trace narrows the dependency beyond a generic VBE failure. All five controller queries INT10/AX=4F00 and both mode-info queries (100h/101h) return AX=004Fh. There is no4F02 mode-set or4F0A request. The original EXECs ./UVCONFIG.EXE (309382bytes) with -gen -p.\settings\; the child returns6 via INT21/AX=4C06 from 0EFE:0163, resuming the parent at 0549:10B7. The parent's4D00 call obtains6; its final4C01 is at 0549:26EB. Missing-file diagnostics for settings/univbe.drv and settings/UVCONFIG.EXE therefore do not mean the packaged root UVCONFIG was absent or never ran. Other unhandled-call/decoder diagnostics remain in the log; none is established as the causal failure by this run.

At41.6272–41.6278M dispatches the child probes sequencer indexesA7/C5 through port3C4 and reads/writes3C5 using helpers 2B5D:0001/2B5D:0028. C5 readback follows writes0/FF; A7 receives87. Current unchanged source masks the index with0Fh, so these accesses address07/05 internally. This is an observed probe sequence plus a source-backed aliasing fact, not proof of a chipset identity or of which conditional makes detection fail. In particular it does not validate the historical optional Trident/InteGraphics selection.

The first remaining dependency is a graphics hardware/register model that the original UniVBE detector can correctly recognize and then drive. The justified next diagnostic is to map the caller and conditional branches around this A7/C5 probe sequence in the original UVCONFIG code, correlating expected readbacks with the captured values. Use that decision to specify a focused extended-register/alias regression before implementing hardware behavior. Do not infer a vendor from successful readback or repair this by raising the advertised VBE version. This task makes no further original attempt.

The matching recovery-run directory/path, EXEC-boundary and mouse contract evidence was reused: all10 recorded source/test hashes still match, with four small logs copied into this run and origin recorded in static-evidence-reuse.json. No broad corpus rerun occurred. test/test-toyvm-vbe.js was not run: the actual original already distinguishes successful VBE queries from later detection rejection, and that mode-rendering test cannot identify the failed chipset branch.

The supervisor caps logs at8000000bytes and stops the original at free disk 2020000000bytes. Actual output is147446bytes; minimum observed free disk was 2108801024bytes, no guard fired, and disk remained above2GB. GNU timeout owned the Node child and reaped it; the Python supervisor actually waited for timeout PID4171289 (exit0, total4.739611s, final wait0.000111s). Its board watcher PID4171288 received TERM and was reaped with exit-15 (wait0.001469s). Node4171290, timeout, watcher and supervisor4171287 were directly absent after completion. Separate publication-watch termination/wait is retained in publication-watch-oswait.json. Shutdown of this task's own processes was explicitly delegated; no coordinator/foreign runtime was stopped.

Recovery: keep the authentic fixture untouched and use the earlier verified archive recovery recipe only if it is actually missing, under new ownership and the stated disk guard. No download occurred here. To inspect this result, read result.json, startup.log, startup.png, identity manifests and wait/ cleanup receipts under scratch/runs/20261101-constructor-original-startup/; regenerate no configuration and do not relaunch the original merely to view the evidence. run.py is the bounded execution recipe, not authorization for a second run. Local box ID/state/expiry: N/A. Browser, boat, EULA, deployment, sound setup and ordinary gameplay remain unperformed.

2026-11-01 authentic UVCONFIG probe branches (static only)

Run 20261101-constructor-uvconfig-probe-branches maps the retained original startup to a concrete generic sequencer false-positive contract and the subsequent exit6 memory gate. A7/C5 rejection alone is not the exit6 cause. No original or emulator was executed, and no source, tests or build changed. The I/O is observed; intermediate registers and branches below are reconstructed from those observations and pinned executable bytes, not an instruction/stack capture. Detector names are the executable's labels, not identified hardware.

Inputs: unchanged authentic28-file closure (63205318bytes), manifest test/toyvm-dos-corpus/files/constructor-demo.json, and its309382-byte UVCONFIG.EXE, SHA256 3a37035f71c213d4dfff0b752415815584f49af1b156afc48f0255b2f993a815. The retained input is 20261101-constructor-original-startup/startup.log (147446bytes), with that run's identity/module provenance unchanged. Read-only source input tools/toyvm/dos.js hashes to eb60e2775196edd80113836830a36482ac50a1d425f3e505d09c4bac68b0f4bc. analysis.json pins these inputs, the manifest's28 files, the ten matching reused contract source/test files, recovery pin record, Python interpreter and GNU objdump. No unchanged contract was rerun. analyze.py uses only Python stdlib byte/trace assertions and /usr/bin/objdump -D -b binary -m i8086 -M intel on entry-aligned slices; it never imports the emulator or executes guest code.

The MZ header is2A00h bytes with2403 relocations. For original segment S and offset O, file offset is 2A00h + 10h*S + O. This child's load segment is0EFEh; add that to original segment operands. This gives the caller chain:

Function Original segment:offset Retained runtime segment:offset
main's detector call 36BE:0872 45BC:0872
initialization/detector wrapper 07EB:DAAA 16E9:DAAA
dispatcher called at DC90 07EB:BFFF 16E9:BFFF
A7/C5 detector, called at C068 07EB:5B28 16E9:5B28
indexed read / write / masked test 1B76:0000 / 0032 / 01CE 2A74:0000 / 0032 / 01CE
raw input / output entry 1C5F:0001 / 0028 2B5D:0001 / 0028

Raw IN is actually at0009 and OUT at0031; the trace's near positions name their entries. Far segment operands are relocation-checked in the executable. Before detector dispatch,149F checks INT10h AX=1A00h VGA presence, including AL=1Ah and display-type checks. The retained original reaches the probes; its VBE query success does not establish a vendor-specific register implementation.

The shared masked test at2A74:01CE saves the register, writes saved & ~mask, reads clear & mask, writes saved | mask, reads set & mask, then restores the saved byte. At0244/0248 it rejects nonzero clear bits; at 024D/0250 it rejects set bits unequal to the mask. It returns AX=1 only when both checks pass, otherwise0. Both C5 calls use mask00FFh. The retained30 I/O rows at41.6272–41.6278M give:

A7 state Saved C5 Write/read clear Write/read set Restored C5 Test AX
00h 00h 00h / 00h FFh / FFh 00h 1
87h 00h 00h / 00h FFh / FFh 00h 1

Detector5B28 is tag0Dh, whose string table says Macronix (pointer table file45770h, string46533h). It saves A7, writes00h, tests C5 at5B95, writes87h, then tests C5 at5BBA. Three PCI searches can accept independently: vendor10D9h with devices0066h,8888h,8625h; each AX != FFFFh jumps to5C65. The three retained INT1Ah AX=B101h/CX=10D9h attempts and later A7 restoration are consistent with all misses. PCI helper2B3B:0112 checks mechanisms1/2 then BIOS presence; an absent PCI service is not itself a generic hardware defect.

Without that override, analysis.json records the computed file offsets and asserts 83 7e fa 00 0f 85 74 02 83 7e f8 00 0f 84 6c 02 at07EB:5C55. These decode to CMP [bp-6],0; JNE 5ED1; CMP [bp-8],0; JE 5ED1. Thus the required pair is 0,1, not1,1. The captured pair takes5C59 to5ED1, restores A7=00h, and returns AX=1 at5EE2 (continue searching). The dispatcher compares AX with1 atC070. A preselected tag0Dh can skip this test, but this trace contains both probes. No vendor identity follows from register aliasing.

The later26 I/O rows at42.9872–42.9877M match a different detector: 16E9:76E8, dispatched atC45F for tag17h. The executable labels17h Hualon (table file45798h, string4659Bh). At7718/772F it tests EDh and EEh with00FFh; 772B/7742 jump to failure782D if either test returns0. Both captured tests save00h, read00h after clearing and FFh after setting, restore00h, hence pass. 7749 stores tag17h. With memory unspecified,77B0 reads E7h; observed00h makes TEST AL,10h at77B8 / JE 77C7 take the256KiB assignment ES:[BX+4]=0100h at77CA. A set bit would assign0200h at77BF. The remaining calls0FFE/04C5 update mode/capability tables; the detector returns AX=2 at7828. The dispatcher comparison atC467 therefore returns that result viaC5C4 rather than trying later detectors.

Wrapper DC99 tests AX=2. DCA4 then compares memory [+4] with0200h (512KiB); DCAA JGE DCB2 is not taken for0100h. DCAC sets AX=6 and DCAF returns viaDE5F. In main,087A decrements the status and the six-word switch at0BA0 maps status6 to089C (status1 instead maps088B, the distinct no-supported-chip error). The status6 path reaches0981 PUSH 6,0983 far call 0000:505B, C-runtime cleanup5001, then0000:015A. At015C–0161, AH=4Ch and AL=argument6 reach INT21h. This matches the retained child exit6 and 0EFE:0163 AX=4C06, then parent resumption0549:10B7. The parent's final “No Graphics hardware” text alone did not distinguish this false detection and memory gate from the child's no-supported-chip path.

The first demonstrated wrong generic emulation contract is unrestricted sequencer writability exposing unsupported extension probes as successful. In the pinned source, the default model is svga='none';3C4 uses seqIndex=value&0Fh,3C5 reads the16-byte array, and vgaSeqWrite unconditionally stores the byte. Thus ED/EE become freely writable0D/0E, while E7 becomes07. The observed A7/C5 aliases07/05 by themselves did not prove causation or Macronix hardware. The later ED/EE probe and explicit memory gate identify the concrete false-positive route. Making C5 obey an invented A7 unlock in the default model would select another detector rather than establish a fix.

Proposed minimal next change: a pure-JS default-Machine contract regression that performs the executable's save/clear/read/set/read/restore00FFh test through3C4/3C5 for EDh and EEh. It must not report both as implemented writable extensions with svga='none'; current code gives1,1 and fails that assertion. The focused fix should make this test pass while preserving defined VGA sequencer behavior (especially map mask and memory mode), register restoration, and existing explicitly selected model contracts. Implementation belongs in the model-specific index-decode/read/write policy in tools/toyvm/dos.js, with normal generated copies afterward. Determine the unimplemented-register read value and legitimate alias policy from the intended default adapter contract; do not replace the16-byte array with256 writable registers or assume FFh reads without that evidence. This task proposes the regression; no new-pass emulator result is claimed. Further valid supported-SVGA detection, at least512KiB, and its real memory/banking behavior remain prerequisites for successful configuration. Neither a VBE version bump, E7 bit fabrication, speculative Trident identity nor hardware/executable substitution follows from this map.

Evidence:15 focused .bin/.disasm pairs,14 pinned instruction-byte checks, switch/label/relocation checks, line-numbered retained trace excerpts and analysis.json in the named run. All28 fixture hashes and exact file closure, all ten reused contract pins, and note whitespace checks pass. No broad tests or build are appropriate for this documentation-only change. Original runtime ownership was never acquired; local box/state/expiry N/A. Only this analysis watcher's shutdown is delegated and directly waited, with cleanup receipts retained. No menu/gameplay/FPS/audio/full27 conclusion is made.

2026-11-01 default sequencer contract repair

Run 20261101-constructor-sequencer-default-contract repairs the demonstrated generic false positive in tools/toyvm/dos.js. The input source is the exact d12f0d1e1 source, SHA256 eb60e2775196edd80113836830a36482ac50a1d425f3e505d09c4bac68b0f4bc. No original game/configurator was executed, and its fixture was untouched.

The existing model declares default svga='none' to be plain VGA. IBM's September 1992 Video Subsystems technical reference, printed page 2-45, figures 2-25/2-26, defines sequencer address bits 2–0 and five data registers, indexes 00h–04h. Higher address bits are reserved, with unspecified readback. The register map does not specify data-port reads at absent indexes 5–7. Consequently zero here is an explicit deterministic ToyVM policy retaining the previous initial read value, not a hardware measurement or an IBM guarantee. No FFh bus-read assumption is made.

The default now decodes three index bits, ignores data writes to absent indexes 5–7 and reads those indexes as zero. Thus EDh/EEh decode to absent 05h/06h instead of writable backing slots 0Dh/0Eh. Defined register storage, map-mask updates, memory-mode transitions and word OUT splitting retain their existing behavior. The shared sixteen-byte storage and explicit selected Trident four-bit decoding, XOR-2 bank writes, version/bank readback and CRTC signature remain unchanged. This adds no chipset identity or memory claim.

test/test-toyvm-vbe.js --sequencer-only exercises actual Machine.portIn/ portOut, using the retained authentic save/clear/read/set/read/restore test with mask FFh. Before the source change both default ED/EE probes report clear=00h, set=FFh, writable=true, restored=00h; the regression exits1. Afterward both report clear=00h, set=00h, writable=false, restored=00h and all69 focused checks pass. Coverage includes omitted/default and explicit none, other absent indexes, defined register bit tests/restoration, index aliases, word OUT, actual planar/rechained pixel preservation, map-mask controls, absent default vendor signature and selected-model negatives.

Full VBE passes102 checks, including its existing guest mode/banked-pixel checks. Existing DAC-default, retrace and Node-only browser-bundle tests pass; no browser was launched. The mandatory flock /tmp/wine-build.lock bash tools/build.sh passes, including generated bundle reproducibility. Both committed ToyVM bundles were regenerated normally.

Evidence retains the old failure, new passes, exact commands/waits, policy reference, retained authentic trace/analysis, source pins and recovery recipe. The95 pinned ToyVM/support/generated/test files are unchanged during tests; the compiler observer records actual module bytes and synthetic COM hashes. VBE uses ToyVM module SHA256 03495d5463de5ee69a82104a0d37d35e7ce9d2662c26f5d06278312c159e080b (343736 bytes); the repair is in host JS, so that module matches the earlier startup's module. The mandatory project build is a separate artifact.

Only this generic contract is repaired. Supported SVGA detection, real memory/banking and the original configuration remain unverified; no playable Constructor claim follows. No E7 bit, model switch, VBE-version change, download, deployment or broad corpus run occurred. Disk remained above2GB; bounded task processes are waited/reaped and no foreign artifact is deleted. Local box ID/state/expiry: N/A; coordinator lifecycle is untouched.

2026-11-01 original startup after the default sequencer repair

Run 20261101-constructor-repaired-default-startup executes the authentic default GAME.EXE once after repair 77a0cea88. UVCONFIG now exits0; the parent reports Problem reading SETSND.CFG (0x0) and exits1 at 0549:26EB. The inspected startup.png is an 80x25 text console with overlapping configurator output, the sound-configuration error and subsequent initialization/SFX messages. No menu, intro, ordinary input or gameplay was reached. This is changed-source startup progress, not validated SVGA support.

The exact GTA recovery worker's author release 124d0bda1, terminal launcher exit0/final and fresh direct absence checks for all seven reported helper PIDs preceded the runtime/source/build claim. Box ID/state/expiry are N/A; own helper/game shutdown was delegated. No source/build changes occurred here.

Before and after the run, all28 authentic manifest files match sizes/SHA256 (63205318bytes), and the directory's complete file set is exactly those28. The fresh actual public DESKTOP_APPS list contains46 entries and excludes Constructor. Source checkpoint 124d0bda1745b544faeafbb6f23f087fac23d6ae has378 unchanged source/support/manifest pins; repaired tools/toyvm/dos.js SHA256 is 5fad1c4e4a64ca69e54c088884b1fe313a61a99300dd6b852095fc579ee0f9c4. All95 accepted regression pins still match. The focused69/fullVBE102 checks, DAC/retrace/bundle checks and mandatory build from the repair run were reused, without rerunning synthetic suites or building.

Native Node v24.18.1 runs the existing default ToyVM tailcall module. The identity-only compiler observer returns the original bytes, recording actual module SHA256 03495d5463de5ee69a82104a0d37d35e7ce9d2662c26f5d06278312c159e080b (343736bytes) and WAT SHA256 29bdf4e754aa46d8e2f083b5bbb8e9fb9c9f2725c4b2503a25b11786234dfbc0. The module matches the earlier run because the repair is host JS. This route does not load the project build/wine-assembly.wasm.

The single command retains default hardware and original executable, with --dispatches=200000000 --seconds=30, supported interrupt/fault/entry tracing, --trace-io=3c4,3c5 --report --text --png=.../startup.png, and GNU timeout90s. No selected Trident/model, VBE bump, E7 fabrication, alternate executable or guest patch was used. Actual log size190150bytes is below the2000000-byte cap.

The relevant comparison is the retained 20261101-constructor-original-startup probe/exit evidence and its accepted static branch mapping. Previously ED/EE both passed writable-register tests, E7 read00h reconstructed256KiB, and UVCONFIG exited6. Here ED's save/clear/read/set/read/restore sequence at 42.9877–42.9880M dispatches reads00h after both00h and FFh writes. There is no EE or E7 probe in the captured port trace, consistent with the mapped early rejection at the first failed Hualon test. No intermediate instruction trace is claimed. The new child executes the same -gen -p.\settings\ command and returns0 via INT21/AX=4C00 at 0EFE:0163; parent 4D00 obtains0. Twelve VBE controller queries and two mode-info queries succeed; there is no4F02 mode-set or4F0A request. The configurator also attempts legacy BIOS modes, as retained in the log; those attempts do not establish correct geometry, banking or visible graphics.

The runner reports creation of settings/UVCONFIG.DAT and settings/univbe.drv and the parent reads the generated driver. These remain in the in-memory overlay and were neither persisted into the fixture nor separately exported. The visible label says InteGraphics IGA168x Internal Clock; this is a configurator label, not proof of an InteGraphics adapter or a valid driver. The accepted branch mapping only explains rejection of the old Hualon path; it does not establish why a later detector accepts. The next graphics dependency is to map that new acceptance against genuine default register/ BIOS contracts and validate its rejection or real implementation before accepting supported SVGA. No second original attempt was made.

The first explicit parent configuration error is the genuinely absent settings/setsnd.cfg; the authentic manifest contains no such file. SETSOUND.BAT still invokes MUSIC/SNDSETUP.EXE from MUSIC and returns to the parent directory. Once graphics behavior is credible, that original ordinary setup workflow is the concrete sound-configuration dependency. No synthetic configuration, sound setup or audio bypass was attempted. Other unhandled-call and six aggregate decoder-giveup sites remain retained; their chronology/causal role in the final exit is not established here.

The parent completes158.3M rounded dispatches and15.81 guest seconds, ending in mode3; these are not FPS/audio measurements. OS runner completion is0, distinct from guest parent exit1 and child exit0. The supervisor actually waits for timeout PID33912 (exit0,5.898180s total), which owns/reaps Node33913. Board watcher33911 is terminated and waited (exit-15); supervisor33637 and all runtime PIDs are directly absent afterward. Runtime disk minimum is 2030686208bytes with the unchanged2020000000-byte stop threshold; a separate independent floor guard also covers preparation/publication above2GB. Own helper cleanup receipts are retained; foreign processes/artifacts and coordinator lifecycle are untouched.

Evidence and recovery recipe are under the named run: result.json, actual screen/log, boundary-excerpts.log, source/fixture/module identities, accepted regression reuse, GTA release gate and separate OS-wait/cleanup receipts. Use retained evidence without relaunching. No download, browser, boat, EULA, deployment, registry/TODOS/STATUS edit, general SVGA, full27, FPS, audio or playability claim follows from this run.

2026-11-01 later InteGraphics detector: static acceptance contract

Run 20261101-constructor-later-detector-contract, input revision 3ce0c5dad7c16a8ee07052597931a3f0e3c269d2, identifies another default register false positive. The later acceptance is a chipset probe accepting ordinary VGA register aliases, not a demonstrated generic VBE fallback. This is reconstruction from authentic bytes, current source and the retained repaired-startup trace; no guest or synthetic suite was executed here.

All28 authentic files (63205318bytes) and exact directory closure match the manifest. UVCONFIG remains309382bytes, SHA256 3a37035f71c213d4dfff0b752415815584f49af1b156afc48f0255b2f993a815; tools/toyvm/dos.js remains 5fad1c4e4a64ca69e54c088884b1fe313a61a99300dd6b852095fc579ee0f9c4. The retained actual ToyVM module is343736bytes, SHA256 03495d5463de5ee69a82104a0d37d35e7ce9d2662c26f5d06278312c159e080b. Its loaded source pins and all95 accepted regression pins still match; reuse the repair's69 focused/102 VBE checks, DAC/retrace/bundle checks and build. Those passes do not cover the newly identified reserved-bit contract.

UVCONFIG file offsets remain 2A00h + 10h*original_segment + offset; the retained child adds0EFEh to relocated segment operands. The detector is original 07EB:B282, retained runtime 16E9:B282. Dispatch atC56A compares the requested tag with24h, C575 callsB282, and C57D/C580 returns any result other than1 viaC5C4. Tag24h's authentic string-table entry is at file457CCh and points to file465E5h, InteGraphics.

Instruction boundaries in original segment07EB Exact predicate or action
B2AC/B2B0 A preselected24h skips probes; otherwise continue to B2B2.
B2B2–B2C0; B2C3/B2C5 Call indexed test 1B76:01CE with port03CEh, index31h, mask003Fh; zero result rejects atB4D8.
B2C9–B2D7; B2DA/B2DC Same test, index32h and mask003Fh; zero rejects atB4D8.
B2E3 Store detected tag24h after both tests pass.
B326/B32C/B32F; B35B/B361/B364; B390/B396/B399 Try PCI vendor10EAh, devices1680h/1682h/1683h. All misses leave subtype unspecified; they do not undo the register acceptance.
B3D3/B3D8; B3DA–B411 If memory is unspecified, read index72h at03CEh and use low two bits:0/1/2 assign1024/2048/4096KiB;3 leaves it unchanged.
B44F/B455 Read index91h at03CEh into descriptor+3Eh.
B4D3 / B4D8 Return AX=2 on acceptance, AX=1 on rejection.

The shared test saves the byte, writes saved & ~3Fh, reads masked clear bits, writes saved | 3Fh, reads masked set bits, and restores the byte. At 1B76:0244/0248 clear bits must equal0; at024D/0250 set bits must equal003Fh. Current dos.js:3340–3354 decodes gcIndex=value&0Fh and stores the entire data byte; dos.js:3080 reads it unchanged. Thus31h/32h alias01h/02h and both tests return1 for any saved byte: clear&3Fh=0, set&3Fh=3Fh, original storage restored. Index72h also aliases02h, and91h aliases01h. These are graphics-controller index aliases, not sequencer ED/EE and not the sound-only ISA port alias in isaAlias.

Observation limits matter: the retained command traces only03C4h/03C5h. It captures repaired ED reads00h/00h, then later three INT1Ah/B101h attempts with CX=10EAh, child exit0 and parent sound error. It does not capture03CEh/03CFh reads, intermediate AX values, descriptor memory or the detector's instruction pointer. The authentic three-device sequence, dispatcher, source-predicted tests and label make this the explained later acceptance path, but its branch-by-branch execution remains inferred. The three PCI attempts are consistent with all misses and an unspecified subtype. A72h read with low bits0 would fabricate a1024KiB descriptor from ordinary GC02h; that particular read value is inferred, not measured. The wrapper's512KiB gate atDCA4/DCAA would accept1024KiB. The actual child exit0 establishes completion without establishing its descriptor contents.

The primary IBM September1992 Video Subsystems reference, printed pages2-76/2-78/2-79, figures2-64/65/67/68, defines a four-bit graphics-controller index and registers00h–08h. GC01h (Enable Set/Reset) and GC02h (Color Compare) have four functional low bits; upper bits must be written zero and have undefined readback. Four-bit index aliasing alone therefore does not prove a defect. The unsupported capability is letting reserved bits4/5 act as independent writable extension state. IBM does not promise zero readback or a particular physical result for this probe.

Next implementable contract: in the default plain-VGA graphics-controller data read/write policy, retain only the defined low nibble of GC01h/GC02h and explicitly choose deterministic zero for their reserved-bit readback. Zero is an emulator policy, not an IBM hardware measurement. Preserve four-bit index decoding, defined plane enable/color compare behavior and register restoration; do not globally discard GC bit masks or selected-model extensions. A focused future Machine-port regression should perform these authentic31h/32h save/clear/set/restore3Fh probes, require rejection while low-nibble tests succeed, and check defined planar write/read behavior. With that policy, set&3Fh can only reach0Fh, so B2C5 rejects this detector. This task writes no such source/test and claims no new emulator pass.

Rejecting this false positive is the next repair, not sufficient SVGA for Constructor. dos.js:835–841 explicitly makes default none plain VGA; there is no implemented InteGraphics model. Existing BIOS VBE1.2 controller responses (dos.js:3988 onward) explain successful query calls, but this accepting detector tests ports and constructs a vendor descriptor without requiring a successful generic VBE mode set. Retained12 controller/two mode-info successes and absent4F02/4F0A cannot validate the generated driver. After truthful rejection, a genuinely supported adapter's detection, memory, bank/window and mode/pixel contracts remain a hardware prerequisite. No particular later detector result or new startup outcome is predicted here.

Sound setup is statically ready for a later ordinary setup assignment once graphics is credible. Authentic SETSOUND.BAT is exactly @cd music, @.\sndsetup.exe, @cd .. (CRLF, no arguments). MUSIC/SETUP.INI contains [PROGRAM] and ConfigFile = ..\settings\setsnd.cfg, resolving from MUSIC to the parent's missing settings/setsnd.cfg. Authentic MUSIC/SNDSETUP.EXE is an LE DOS program with setup.ini at file22004h and ConfigFile at22A74h. Its bounded unrelocated object1 slice at relative15C0h selects PROGRAM;162Eh–1648h passes ConfigFile and destination6CD8h to the lookup routine. The save slice at2176h/2198h/21E1h reuses that destination as a path and includes distinct create/open/write-error branches. These are static file/object offsets, not observed runtime addresses or proof that saving succeeded. The INI also names digital/MIDI tests and supplied HMI support files; test.wav is not in the authentic28-file closure, so digital test-sample availability remains an explicit later setup limitation. No sound setup, fabricated SETSND.CFG, alternate hardware/executable or original rerun occurred.

Evidence: analysis.json pins authentic inputs, actual retained module and accepted tests;13 instruction-byte checks, relocation checks, six bounded disassembly slices, source excerpts and line-numbered retained interrupt/port excerpts distinguish observed facts from inferred branches. sound-inputs.txt retains the authentic invocation/INI. Recovery is static: rerun only analyze.py against those unchanged files, never the original run recipe. No build/suite rerun was appropriate for this note-only change. Runtime/ source/build ownership was never acquired; GTA's lane is untouched. Box ID, state and expiry are N/A; guest exit is N/A for this task (retained child0, parent1). Resource failure: the independent guard recorded1975025664bytes free, below the required2000000000-byte floor, after15 samples. Its watcher was terminated and actually waited (exit-15); the guard exited0 and both PIDs43755/43757 were directly absent. This violation was discovered during finalization, after the static analysis above had completed; the guard's early exit did not interrupt the analysis driver. No sustained-headroom success is claimed. Further analysis stopped on discovery, without retry, cleanup of foreign files or a runtime launch. The guard needs a live driver abort signal in a future assignment. Evidence/publication only follows; no RAM gate, browser/boat or gameplay/audio claim is involved.

2026-11-01 GC defined-bit repair: tested WIP, build aborted

Run 20261101-constructor-gc-defined-bits-repair, assigned input c2105cd0d10645199ec59dda8ece6d5a1dd1209b, prepared the next default-VGA contract above. The repair remains uncommitted WIP; the mandatory build did not finish. Only this investigation note is published. WIP paths are tools/toyvm/dos.js, test/test-toyvm-vbe.js, and the two existing generated docs/dos-corpus/live/toyvm{,-jit}-bundle.js files.

The proposed data-write policy stores value & 0Fh in GC01h/02h only when the selected model is plain VGA (none, including the default). Reads then return zero for reserved bits. This is deterministic emulator policy; IBM September1992 pages2-76/2-78/2-79 specify undefined reserved-bit readback, not guaranteed physical zero. Four-bit index decoding, other GC data policies and explicit Trident behavior remain unchanged. Both bundles were regenerated; the build's reproducibility check accepted them before the later abort.

The real Machine-port regression performs the authentic GC31h/32h mask3Fh save/clear/read/set/read/restore sequence with initial low nibbles0/5/A/F. On the unchanged old source it exits1 with20 failed and54 passed assertions; with the repair it exits0 with74 passed assertions. Both default and explicit none return clear00h/set0Fh and restore the saved low nibble, rejecting the extension predicate while preserving low-nibble functionality. Word OUT, index aliases, other GC registers and the selected-model compatibility checks are included. This is synthetic port evidence, not an observed original UVCONFIG branch execution or a qualified SVGA model.

The affected combined VBE suite passes178 assertions, including the existing 69 sequencer and33 VBE assertions and two new guest-planar checks. A synthetic COM configures ports and executes real A000 planar memory reads/writes to check enable-set/reset planes, matching/nonmatching color comparison and restoration to CPU-data writes. Its SHA256 is d760de7e6c38f2755bf18fdffa4cc054d61e25d667014fb6aa3abb1183222d5a; the actual emitted module remains 03495d5463de5ee69a82104a0d37d35e7ce9d2662c26f5d06278312c159e080b. An initial fixture omitted Color Don't Care and failed one assertion; the corrected fixture explicitly compares all four planes. Both logs are retained. No DAC/retrace/bundle execution suite or original was run in this assignment.

First blocking prerequisite: flock /tmp/wine-build.lock bash tools/build.sh was actively terminated when free /home disk reached1988030464bytes, below the retained stricter2020000000-byte stop threshold and required2GB floor. The command supervisor sent TERM to build process group62297, actually waited for its driver (exit-15), and killed remaining group children. Direct inspection found no remaining group members. The persistent independent monitor also latched that same minimum and exited0, actually waited; no sustained-headroom success is claimed. Unlike the previous guard failure, this breach interrupted the active task driver. The build reached combined WAT generation, so it cannot be reported as a completed compile/build. Implementation and tests stopped without retry, floor reduction or foreign cleanup. The separately prepared injected-breach self-test was not executed.

Evidence includes baseline hashes, actual test/module/fixture identities, old/new/fixed logs, guarded-command wait receipts and the latched abort. Restore sustained headroom and assign a fresh continuation before resuming; inspect the preserved WIP, run affected remaining checks and the mandatory locked build, and commit these code paths only after all required checks pass. The current passing synthetic checks do not establish playable Constructor. A genuinely supported adapter's truthful detection, memory capacity, bank/window access and mode/pixel behavior remain prerequisites before any separately assigned original or authentic sound-setup run. No hardware switch, fabricated memory/E7/vendor identity or setup bypass was introduced. Box ID, state and expiry are N/A; no test box or original game was started.

2026-11-01 GC final validation: checks pass, runner dependency blocks publication

Run 20261101-constructor-gc-final-validation resumed after the GTA checkpoint worker's final result, author release and direct absence of its nine reported processes. All four GC source/test/bundle SHA256 pins still matched the preserved repair. The accepted source-suite result remains178 passing assertions; that unchanged suite was not repeated.

The remaining DAC-default and retrace suites both exit0. Each preserved bundle also passes the74 GC port assertions through its own bundled Machine, and executes the existing synthetic COM bundle test to print HI and exit. Both embed byte-identical repaired dos.js. These are Node-hosted bundle execution checks, not a browser or original-game run. The injected disk-guard self-test actively terminated its driver and child, actually waited and verified their absence. Real validation commands stayed above2338738176bytes available; the unchanged stop threshold is2020000000bytes and floor2000000000bytes.

Concrete unresolved dependency: GTA left its tested checkpoint repair in uncommitted tools/toyvm/run-dos.js and test/test-toyvm-runner-input.js. Comparing every embedded module in each preserved bundle against current source finds exactly one mismatch: tools/toyvm/run-dos.js. The preceding GTA mandatory locked build failed at that reproducibility gate (retained in scratch/runs/20261101-gta1-checkpoint-capture-fix/build.log). No unchanged build retry was made here. Regeneration would incorporate that foreign runner repair; publishing those bundle changes without its source would leave main inconsistent. This task explicitly excludes the runner/test WIP, so the four GC code/test/bundle paths remain unchanged and uncommitted. Only this note is published; the mandatory build still has no passing result.

The next continuation must reconcile ownership/publication of the runner and generated bundles, then pass flock /tmp/wine-build.lock bash tools/build.sh with at least2150000000bytes entry headroom and the active disk guard before committing code. Reserved-bit zero readback remains deterministic emulator policy, not an IBM physical guarantee. Supported SVGA detection, capacity, banking, mode pixels and Constructor gameplay remain unproved. No original, browser, box, download or deployment was started; box ID/state/expiry are N/A.

Combined GC/checkpoint implementation outcome, 2026-11-01

Run 20261101-toyvm-gc-checkpoint-coherent-build adopts the six released source/test/bundle paths together and resolves the runner dependency above. Both bundles are regenerated from the repaired GC and checkpoint runner; all embedded modules match source after shebang normalization. Each passes 74 GC assertions, real HI COM execution, once-only checkpoint keyboard input, no-input equivalence, session bounds, stop, deadline and terminal checks. Exact source/test hashes retain the accepted 178-pass source VBE result, DAC/retrace PASS and native checkpoint/keyboard PASS from the preceding runs. flock /tmp/wine-build.lock bash tools/build.sh now passes in full, including bundle reproducibility. Evidence and recovery instructions are in this run.

GC01/02 store only their defined low nibble for default/none SVGA; reserved-bit zero readback is emulator policy. IBM reserved reads are undefined, so this does not establish physical zero readback. Selected Trident behavior and other GC registers retain their existing policy. No original was executed here; supported SVGA detection/capacity/banking/mode pixels and gameplay remain unproved. Build entry was2335780864bytes available and its minimum2335588352, above the active2020000000-byte stop and2000000000-byte floor, with no RAM gate. Box ID/state/expiry are N/A; runtime is released after publication for the next authorized original test.

Original default startup after GC repair, 2026-11-01

Run 20261101-constructor-post-gc-startup makes one original default launch after published GC repair a416f0142, on 37791d8f1 (the subsequently published checkpoint/stdin repair). Graphics detection now rejects the default VGA: UVCONFIG exits1 and GAME.EXE exits1. The inspected startup.png shows “No supported SuperVGA chip detected!” and “Error = No Graphics hardware detected by UniVBE! (0x0)”. This is the first changed-source original failure; no sound setup, menu, input or gameplay was reached. The earlier apparent InteGraphics acceptance and missing-SETSND.CFG boundary no longer occur.

The checkpoint worker's launcher exit0, final/result, published author release and fresh direct absence of all19 reported helper PIDs preceded acquisition of native runtime ownership. No emulator source, test or build was edited or run here. Its accepted focused/keyboard/bundle/parity/build results are retained in dependency.json; the GC host source is byte-identical to a416f0142 (SHA256 6b8d9420c7d6fd2653838d1720d3acfe16f499e059bee13d9e44dfc8cc00dd85).

All28 authentic files, totaling63205318bytes, match the retained manifest before and after execution, with exactly that file closure. GAME.EXE remains 994161bytes, SHA256 d0ac2f05b9a813dc723f170f7418fbfd9af269abcbce8905062c3b63f27dc587. The manifest itself matches the retained archive-verification record; that record pins the32062553-byte archive to 5d764f6d2f536d6aae6fd18122bc9bf2be20e84f101cbcdeafb1bcbd20eeed18. The archive has been removed locally, so this is reuse of its recorded verification, not a fresh archive hash or download. All378 runtime/support/ manifest source pins remain unchanged during execution. The identity-only compiler observer returns the original bytes and records actual module SHA256 03495d5463de5ee69a82104a0d37d35e7ce9d2662c26f5d06278312c159e080b (343736bytes), matching the retained module; the GC repair is host JS.

The sole command uses the original GAME.EXE, default hardware, no command tail,200M dispatch/30 guest-second limits and an outer90-second timeout. Existing interrupt/fault/entry and3C4/3C5/3CE/3CF port diagnostics, text/report and PNG capture are enabled. No DEMO route, model/vendor/memory/E7 override, forced configuration or binary patch is used. The CLI reports2.86seconds, 43.7M rounded dispatches and4.36 guest seconds; these are not gameplay or FPS measurements. The timeout/Node host completes normally (actually waited, code0), distinct from both guest failures: child INT21/4C01 at 0EFE:0163, parent 4D00 obtains1, and final 4C01 at 0549:26EB.

At43.2197–43.2199M dispatches, the captured GC31h save/clear/set/restore sequence reads00h, writes00h and reads00h, writes3Fh and reads0Fh, then restores00h. No GC32h probe follows. This observed port behavior matches the authentic detector's mapped early rejection predicate; no instruction pointer or descriptor capture is claimed. Five VBE controller and two mode-info queries succeed, but there is no4F02 mode set. The text console and final mode3 do not establish supported SVGA, capacity or banking.

The concrete next work is a separately owned complete adapter contract that the original detector genuinely supports: derive its identity/register and capacity checks from authentic UVCONFIG code, implement real bank/window access and mode behavior, and verify bank-crossing pixels before another original startup. Detector acceptance alone cannot qualify an adapter. Default plain VGA rejection must remain truthful.

The conditional sound follow-up was not triggered. Original SETSOUND.BAT and MUSIC/SETUP.INI contracts are retained in sound-inputs.txt; no SETSND.CFG was fabricated. Correction to the older missing-test.wav note: the retained archive manifest and current authentic closure contain MUSIC/TEST.WAV (67024bytes, SHA256 d972ec69430baa17b93b10bf13871ff913de41963257e004eea3249f5e506852). Runtime sample availability and audio playback remain untested limitations; no asset was synthesized or setup/audio success claimed.

Evidence, identity checks, command/wait receipts, boundary excerpts and recovery instructions are in the named scratch run. Original-command free disk minimum was2312757248bytes, above the active2020000000-byte abort and 2000000000-byte floor; no RAM gate applies. Local box ID/state/expiry are N/A; no box lifecycle action occurred. Own helpers are actually waited and their direct absence checked before final release/publication handoff.

2026-11-01: bounded optional Trident SR0B/SR0D access repair

Run 20261101-constructor-trident-register-contract follows the released GTA input task 52165cda8. This is a synthetic port-contract repair, with zero original GAME/UVCONFIG executions and no gameplay claim. It does not establish why the earlier optional Trident run selected InteGraphics. Default and explicitly nonselected VGA still reject extended sequencer storage; merely selecting svga=trident is not supported-adapter evidence.

The missing contract is independent old/new SR0D control storage: writing sequencer data SR0B selects old controls; reading its data selects new controls. Selecting index 0B alone has no effect, and repeated reads select new controls rather than toggling. The original register descriptions are Richard Ferraro, Programmer's Guide to the EGA, VGA, and Super VGA Cards, 3rd edition, section 27.5.1, pp. 1366–1368, figures 27.6/27.8 and listings 27.3/27.4 (text reference). The hardware-driver primary source also explicitly enters new mode by reading SR0B: Linux drivers/video/fbdev/tridentfb.c, enable_mmio and disable_mmio (upstream driver). These are a programming reference and an actual hardware driver, not a newly obtained manufacturer datasheet or an inferred UVCONFIG trace.

The optional model now stores old SR0D separately from its existing new SR0D and applies SR0B's access side effects. SR0B data writes cannot invent a version; reads return zero as an absent-identity policy, not a qualified TVGA8900 version. The test uses valid old/new control-bit values and checks data access versus index selection, repeated selection, independent machine state, unchanged default/nonselected negatives, and persistence at both endpoints of existing video banks 0 and 1. Against the saved old source, the exact final focused test fails three assertions: two independent SR0D readbacks and prevention of writable chip identity. The repaired source and both regenerated bundles pass all 13 checks each. Full VBE/sequencer/GC, default DAC and retrace regressions pass. Both bundles match every embedded source module (44 interpreter, 43 JIT) and each executes the existing real COM test to print HI. Commands, old-fail/new-pass logs, bundle parity and actual wait receipts are retained in the named scratch run.

This deliberately does not qualify a supported Trident adapter. Only SR0D's independent storage is repaired, not its clock/mode effects or the full old/new sequencer register bank. SR0B's real chip version is absent; the pre-existing TVGA8900 ROM string, synthetic SR0E high-bit readback and CRTC1F XOR signature are not identity/capacity validation. SR0E still uses the existing 64 KiB XOR-2 bank implementation even after old selection; real old-mode 128 KiB windows/control behavior are missing. The generic 1 MiB backing has not been tied to a real chip's installed memory, CRTC capacity programming or wraparound. Four-bit sequencer-index aliases and extended timing, clocks, DAC and mode-pixel behavior remain unqualified. A complete advertised identity/capacity/banking/mode-pixel contract remains necessary before any original retry. The truthful default VGA rejection from cf7ba641a remains the last original Constructor result.

No box was allocated (ID/state/expiry N/A); coordinator lifecycle ownership is unchanged. Evidence remains under scratch/runs/, outside git. No assets, configuration, binary patches, deployment or worker launch accompanied this repair. The mandatory locked build and final publication receipt are recorded in the run evidence.

2026-11-01: old-mode SR0E/window research stopped at unresolved semantics

Run 20261101-constructor-trident-old-window starts from published 995c90daf. No hardware repair is claimed: the references inspected do not justify a coherent old/new CPU bank/window contract. Source, tests and both bundles are unchanged by this task; no synthetic signature or chip identity was added. No GAME/UVCONFIG execution occurred. The GTA batch task's 4f8051361 final/result, terminal-wait evidence and exact author release were read, and all31 reported helper PIDs were directly absent before owning this note. No foreign WIP was adopted.

The actual SVGAlib TVGA8900 driver, saveregs/setregs, saves/restores independent old/new SR0E alongside SR0D. Its setpage explicitly selects new controls and writes the page XOR2; saved new SR0E readback is also XOR2 before restoration. The Xorg bank driver uses identical SR0E values for its read, write and combined bank setters. Neither supplies an old-mode CPU banking routine or a memory test across SR0B selection. These sources support register independence and the new register write/readback convention, but do not settle the old window.

The unresolved behavior is concrete:

The Trident TVGA8900D manufacturer datasheet was retained and visually surveyed as a23-page scan, with page5 read at full size. It describes the bus/memory interface, but supplies no SR0D/SR0E programming definition that resolves these questions. Its capacity overview does not qualify ToyVM's installed memory. A later GUI9420 register-document URL was inaccessible (HTTP401). MAME was inspected as a secondary emulator only; importing its old-bank mask and new-window offset alias would not be hardware validation. Firsthand CRT Terminator conformance results also report differing old/new banking on two8800CS boards; that evidence does not define a8900 contract.

The next requirement is a manufacturer programming reference or reproducible real-chip memory trace resolving the bank stride, paging/decode prerequisites, SR0B transition effects and upper-window behavior for one explicit pixel configuration. Until then, boundary/persistence expectations would encode assumptions. Under this task's reference-blocker stop condition, no new old-fail/new-pass tests, bundle regeneration, VBE/DAC/retrace runs or locked build were performed. Earlier passing checks remain earlier evidence, not validation of a repaired old window. Identity, installed capacity, mode pixels and authentic UVCONFIG support remain unqualified; default VGA's last original rejection remains the applicable result.

Retained sources, hashes, failed retrievals, static baseline pins and detailed questions are in the named scratch run. Box ID/state/expiry are N/A; no box lifecycle action or emulator runtime acquisition occurred.