Lords2 installer memory-check path
Task: NEW-GAME-LORDS2-DEMO-20261015. Evidence: scratch/runs/20261106-lords2-memory-check-path/. Guest code, media and configuration are unchanged. This investigation follows the API-only observation; KERNEL.131 is GetDOSEnvironment. The independently reproduced GetFreeSpace undercount has no established consuming Lords2 trace.
Preparation and retained evidence
The predecessor's warning capture was personally reviewed. It says “You have [not tested] memory ([not tested])”, requires 7700 KB physical memory, and is installer evidence, not gameplay. The retained 20261105-lords2-stack-pointer-writer/live-write-history.json contains 503 writes but only 36 distinct executed writer blocks; writer and previous-block code together give 61 contexts. These concern the earlier stack-cell repair and do not reach this result writer. The retained retirement readout likewise has selected teardown evidence rather than general executed-block coverage. Details and provenance are in retained-executed-block-review.json.
Read-only disassembly of the retained _SETUP test executable (20261105-lords2-recovered-original-gameplay/generated-setup.exe, SHA-256 7dbf696d4f08464ca1e187686368fe89c5b71b3162c094f463d0e5ec4eaca567) locates two candidate memory methods: segment 4 offset 3F04 calls KERNEL.169 GetFreeSpace; 3F4E reads CMOS through ports 70h/71h. The memory aggregate constructor is 4:3FB8; its subobjects use keys PhysicalMem and MemKB. These are static candidates, not evidence that either query ran. An apparent CD 31 at 3:31F6 is inside a far CALL operand, not an executed INT 31h instruction.
The changed run-local observer uses the shipped set_win16_trace and set_trace_esp diagnostics. It forwards existing host imports, retains first executed-block contexts and code bytes, and bounds repeated state capture around the known child test/helper code. It records all selected profile-string calls and outputs, LoadString, GetVersion, WinExec and warning formatting, including their caller rings. The generated qualification executes ordinary INT 31h, INT 2Fh, INT 21h, a near helper and imported memory APIs; armed/disabled architectural state and all 512 MiB of memory must match on both shipped modules. No production instrumentation is added.
Aranna owns local runtime/source/build during preparation. Generated runtime qualification and the authorized single changed original trace wait for its explicit release and verified cleanup. Pure parser controls and in-place exact 181-file media verification pass. No original has launched in this preparation phase.
Executed consumer and repair
Aranna explicitly released the runtime at 23:56:53Z; independent PID, port, profile and three-lock checks passed before use. Generated observer qualification passed on both standard and compatibility modules: real INT31/INT2F/INT21, near helper, imported APIs and CMOS instructions; armed/disabled registers, lazy flags and complete 512 MiB memory hashes matched, including bounded-buffer saturation. The first generated attempt timed out in load_ne: the run-local fixture builder used an aliased Buffer slice and cleared its own relocation targets. Replacing that with an independent Buffer copy fixed the generated fixture. Neither failed attempt launched the original; both diagnostic logs are retained.
Exactly one changed registered original ran, 2026-11-07 00:03:56Z–00:05:42Z, normal exit 0. Its source was 2c540540bf9ba94954e79e92ae2d2a142eaad2d2; standard module SHA-256 6cd858302895de91df5af17e3a3742281e31f369ddb6528d7b96252b3423caac. Exact existing 181-file media verification passed. Startup, Install, System Test Yes, Wave Yes and CD-ROM OK captures were personally reviewed. Wave Yes is an ordinary dialog answer, not proof of working audio. after-cdrom-ok.png reproduces the Memory Test warning; no gameplay was reached.
memory-warning.json contains the executed evidence. The child trace has 3810 unique blocks, no block overflow, record drops, focused drops or observer errors. The causal path is:
- PhysicalMem object
001F9242, vtable001BD374, holds required KiB 7700 and measured KiB 0. Its real method4:3F4Eexecutes at block 7061552. - Records 166–173 show ordinary OUT 70h / IN 71h pairs selecting 15h, 16h, 18h, 17h, with all four reads returning zero. Registers in the same method show BL/BH becoming zero and AX becoming zero. This is the first demonstrated incorrect emulated operation, independent of Win16 API tracing.
- Executed instructions at
4:3F7B–3F87add base and extended KiB, store the zero result locally, and call virtual setter+38h. At4:38CE, block 7061558, AX=DX=0 and the actual Pascal arguments are object0087:9242, value0000:0000. - Profile write record 174 consumes those results: section
Config, keyPhysicalMem, string value0, fileSIERRA.INI. This is ordinary guest output; the observer does not supply it. - The formatter
4:359Creads the same object. Its zero branch4:35BBexecutes at block 7062013 and copies the default string. The personally reviewed warning displays[not tested]for physical memory. The separate MemKB object001F9272has both measurement and requirement zero and uses the same default branch. No4:3F04GetFreeSpace call executes in this path; the earlier generated undercount is not this physical-memory failure. No new INT31 hypothesis is needed.
The observer's GetPrivateProfileString annotation has an off-by-one field-index error for section/key/default pointers (its raw arguments and returned buffer are intact). Those annotated strings are excluded from the causal claim. The actual setter arguments, decoded WritePrivateProfileString record, object snapshots, executed blocks and CMOS records above do not depend on that annotation. No second original was run to repair diagnostics.
IBM's AT Technical Reference, September 1985, System Board CMOS description, defines 15h/16h as binary low/high base-memory KiB and 17h/18h as low/high extended-memory KiB above 1 MiB. The host's existing CMOS constructor only initialized RTC registers A/B, leaving the memory configuration at zero. The repair initializes these existing writable RAM bytes to 640 KiB base (0280h) and 63 MiB extended (FC00h), consistent with the legacy GlobalMemoryStatus 64 MiB machine. The 384 KiB adapter hole is excluded from the sum; this does not report the emulator-private 512 MiB wasm buffer as guest RAM. Calendar BCD mode does not affect these binary configuration bytes. Device ownership, shared process state, writes, callbacks and lifecycle remain unchanged. The required callback inventory hash receives a narrow lifetime review, preserving its pending ownership status and all gates.
test/test-cmos-memory.js executes an ordinary generated two-segment NE with CLI, byte IN/OUT, BL/BH and AH/AL assembly, STI, ADD/ADC and guest stores. Before the repair it fails at base memory 0 !== 640; after repair both module variants produce base 640, extended 64512 and sum 65152 without overflow. Additional checks preserve writable RAM, NMI masking, RTC mode independence and process isolation. Existing RTC host, synchronous broker and actual Worker controls pass. The required full build and final verification are recorded in the run logs.
The source repair is demonstrated by the generated consumer and the original's exact failing operation. No post-repair original installation or gameplay progress is claimed. The one authorized original was already closed before changing host initialization; it cannot acquire new constructor defaults. Next is an ordinary post-repair installer run, with fresh authorization for that additional original, to establish visible progress and discover any later blocker. No guest/media/config patch, forced result, branch override or observer guard relaxation was used. No boxes were assigned or created.
Final validation: flock /tmp/wine-build.lock bash tools/build.sh passed (build-reviewed.log), both freshly built module variants passed the new regression, RTC host/broker/production import sharing and real OS Worker checks passed, and test/test-x86-ops.js passed 293/293 (focused-post-build.log). Wasm hashes stayed identical because the repair is host initialization. Post-run exact media verification passed. Runner, supervisor, server and Chrome are absent; port 36924 refuses connections, the owned profile is absent and all runtime/build locks were independently acquired/released (cleanup-final.json). No safety flag occurred. The run retains source SHA plus binary patches and unique evidence, with no source/media archive. Runtime ownership is handed back; no assigned boxes require lifecycle action.