Lords2 post-repair installer run

Task: NEW-GAME-LORDS2-DEMO-20261015. Evidence: scratch/runs/20261107-lords2-repaired-cmos-install/. This follows the executed memory-check path and CMOS repair, published as e5c075a59ea84bcb3abc0334c6e3986ebc96a96b.

Preparation verified the exact registered 181-file original fixture in place: 14,382,323 bytes, inventory SHA-256 f73d8da675e08d1a5460e380e1b314a4bfe7cc59031ec1202b8f11560f929afc, browser manifest SHA-256 a1749143a2de176c4f3fdf314337a1767f6101eed4a88b6d5b0dee4d59012f8e. No source/media archive or guest modification is used. Git SHA and relevant binary diff preserve source identity.

The existing bounded path observer is reused unchanged, including its documented GetPrivateProfileString annotation limitation. Its generated armed/disabled architectural and full-memory equivalence controls must pass on both shipped modules after explicit Aranna runtime release and independent cleanup verification. The existing focused CMOS regression must also pass before the single authorized changed original. Preparation alone establishes no post-repair installation or gameplay progress.

Execution and actual progress

Aranna explicitly released runtime/source/build at 00:43:29Z. Independent checks found all seven recorded session process trees absent, their ports refusing connections, their profiles absent, holder PID 1110100 absent, and all four runtime/build locks independently available (handoff-cleanup.json). Both standard and compatibility modules passed the existing CMOS regression and the unchanged observer's generated interrupt/helper/API/CMOS controls, including bounded saturation, architectural equality and complete 512 MiB memory equality. No new observer was used.

Exactly one changed original ran from 2026-11-07 00:43:58Z to 00:46:26Z, normal harness/Chrome exit 0. Source commit was 2e456d08de4f5eb6e11be89cfc9923e860431fb5, with an empty relevant tracked source diff. Actual HTTP receipts match the repaired lib/host-imports.js SHA-256 29fc6ffca2b03e869a5c40969a0c0ecfee4b5489a372d9974bf64b3a1772febf. Both instantiated modules identify standard wasm SHA-256 6cd858302895de91df5af17e3a3742281e31f369ddb6528d7b96252b3423caac; compatibility prerequisite used 41b0c704c91ba6fc7e4a6645d9421bf7fd3a23139b0b538b576e1e1b85527dd6. Tested module copies are retained. The environment was local headless Chrome 151, 1000×800 viewport, cooperative child execution, ?debug&no-threads.

Personally reviewed captures show Main Menu, Install, System Test Yes, Wave Yes, CD-ROM Detect OK, then removal of the installer child windows. Wave Yes reproduces the ordinary installer route and is not proof of working audio. The actual memory consumer now reads CMOS indices 15h/16h/18h/17h as 80h/02h/FCh/00h, and writes Config/PhysicalMem=65152 to SIERRA.INI (records 166–175). This confirms the repair in the original. No post-repair physical-memory warning was displayed, but the system-test results did not finish displaying either. No installation, genuine installed state, installed-game launch or gameplay is established.

First remaining blocker: PIT channel 2 calibration

first-blocker.json retains the consuming code, contexts, registers and terminal ring from after-cdrom-live.json. The original child has 3911 unique executed blocks, 180 selected records, no observer errors, block overflow, focused drops or record drops. Calls and returns balance. No GetFreeSpace consumer is observed.

The next executed consumer is original _SETUP segment 1:9DF6, a CPU-speed calibration routine. Its helper 1:9E5D writes control B4h to port 43h, count FFFFh to 42h, and changes port 61h bit 0 around its instruction loops. At 1:A0AB–A0B3, it reads port 42h twice, assembles AX and negates AX. Intel's 8254 datasheet, order 231164-005, control-word table on page 6, specifies B4h as counter 2, low-byte/high-byte access, binary mode 2. Its counter and gating semantics require a running counter rather than floating-bus bytes.

The tested emulator's $th_port_io in src/06b-core-handlers.wat initializes IN results to all ones and has no port 42h branch. Channel 2 control/count writes are ignored; port 61h always reads zero and its writes are ignored. Thus the two reads in this executed path return FFh/FFh, assembling FFFFh and negating to 1. These port values follow directly from the tested source handler; the existing observer did not separately instrument port 42h. The captured helper returns independently agree: entry to 1:9E17 has AX=1; entry to 1:9E21 has AX=BX=1. Its SUB AX,BX; STOSW therefore stores zero. The five-sample accumulator reaches the unique final executed block 1:9E3A with AX=DX=0, BX=1, SI=DI=001Ah, SS=DS=ES=0087h, CS=0047h, ESP=001F8A6Ah.

Inside that block, MOV BX,5; DIV BX averages zero, then MOV BX,AX makes BX zero. MOV AX,FA00h; MOV CX,04A9h; XOR DX,DX; MUL CX produces DX:AX=048D:0A00. The next instruction, 1:9E4B DIV BX, therefore has a zero divisor. The existing $th_div16 raises C0000094 for this condition. The console directly records child retirement Thread 1 exited (postYield=2) prev_eip=0x179e3a esp=0x1f8a6a, and captures show the resulting empty desktop.

The exact 9E4B instruction and exception code are derived from the captured final block's entry registers and straight-line code, not a separately captured fault-time EIP/exception record. Repeated-block contexts retain first visits; terminal ring and unique final-block context remain available. The precise source defect is absent PIT channel 2 state/read/count/gate behavior; the DIV zero check should remain strict. A repair needs a focused generated calibration consumer and timer semantics, rather than a forced CPU-speed result or guest branch change. No production repair was made in this run, and there was no second original or diagnostic rearm.

Validation and handoff

Exact 181-file media verification passed before and after execution. Of 213 HTTP receipts, 203 match local bytes exactly. The two index responses match the reconstructed normal dev-server injection; eight retained 404s are build-info and CARDS filename probes, also seen in the predecessor. Repaired host imports and the actual wasm response match exactly (http-review.json). The GetPrivateProfileString annotation error is excluded from the causal argument; CMOS results, decoded WritePrivateProfileString arguments, executed blocks and registers do not depend on it.

The original supervisor, runner, server and Chrome are absent; ports 36927 and 42737 refuse connections and the owned profile is absent. All four locks were independently acquired/released, free space exceeded 2 GiB, and explicit runtime/source/build RELEASE was appended at 00:46:47Z (cleanup-runtime.json). No boxes were assigned or created. No source/media archives, guest patches, forced results, guard relaxations, unchanged relaunches or safety flags occurred. Evidence review and publication finish offline; result.json is published last. Next work is a focused PIT channel 2 implementation/regression, followed by a separately authorized original run to establish installation and gameplay.