Lords II prevalidated setup observation — 2026-11-05

One fresh original observation reached Install, System Test Yes and Wave No through production-validated ordinary canvas input. Wave No reproduced the Win16 indirect far-call trap and retained the actual operand, selector, guest registers and live segment/module ownership. Installation and gameplay remain unverified. No causal emulator repair was demonstrated or committed.

Evidence: scratch/runs/20261105-lords2-prevalidated-setup/.

Identity and preparation

The required final releases were Stronghold1425 at 14:49:14 UTC, formats1430 at 15:11:01 and ReturnFire1435 at 15:27:01. Independent inspection then found 175 predecessor PIDs absent, 22 ports refusing connections, 14 profiles absent, four free locks and no heavy runtime. No shared edit, executable check, build, browser or module pin preceded those releases.

Tested source was main 7e3bf7788b229993266bb30b0f1d7b813163e04a. The browser compiled the standard module, 2,173,299 bytes, SHA-256 10462bd9693cf4f51d382d33767674b736b84e74346f328306d59e35c5e60031. Startup and final readouts agree on its identity. The cooperative backend ran the faulting child T1 in process 1000. Chrome was Linux x86-64 version 151; viewport 900 by 720. The local runtime held the existing local/heavy/browser flock locks.

The published 6a0ce7e7198 diagnostic, host collector and six-case regression sources were unchanged. Their retained standard/compat NE, collector lifetime, retirement/interleaving and full-build evidence was reused. Intervening Stronghold changes concern Win32 IO retry ownership; formats changes concern GPU mip gradients. The current canonical module also matches ReturnFire1435's passing full build. No fresh full build or generated NE run was needed for this observation; no emulator source changed here.

Cheap checks passed: helper JavaScript syntax, production queue-validator self-test, original SETUP MZ/NE header, all 181 canonical media inventory hashes (14,382,323 bytes), and browser manifest SHA-256 a1749143a2de176c4f3fdf314337a1767f6101eed4a88b6d5b0dee4d59012f8e. Media remained at scratch/game-media/lords2-demo-20261105/original; no media or source archive was copied. SETUP.EXE SHA-256 was c9844241cf6594050cb4cdde5b04d4f994dfa6da83a3c5538dab16e792aa49b7.

Live failure

The caller started at 15:28:00.420 UTC, with a 540-second outer bound and the production queue's 420-second limit. All three clicks and the final quit were prevalidated by the actual production queue validator. Each click was derived from the current personally inspected screenshot and live canvas rectangle; the retained review records and harness output include fractions and geometry. The screenshots show Main Menu, System Test, Wave question, and the desktop after the child crash. No pixel-style continuation command was used.

At 15:29:52.365 UTC the diagnostic captured:

Item Live value
Calling module / segment _SETUP, module 15, NE segment 4
CS / base 005f / 001a0000
DS = SS / base 0087 / 001f0000
Block / return 001aa26c / 001aa297
ES / BX after LES 005f / 869c
Operand address 001a8700
Operand bytes 00 eb 29 8b
Target selector:offset 8b29:eb00
Pre-push ESP 001f8554

The live call bytes include 26 c4 1c (LES BX, ES:[SI]) followed by 26 ff 5f 64 (CALL FAR ES:[BX+64h]). The captured stack object at 001f88a4 begins 9c 86 5f 00; LES therefore produces 005f:869c. Adding displacement 64h yields exactly the captured operand address. The target selector has no mapped segment. The preserved CS rejection traps, and the crash record shows ESP 001f8550 after the existing four-byte return-frame push. Guest bytes and CS checks were not altered.

A read-only comparison with the earlier retained generated _SETUP image SHA-256 7dbf696d4f08464ca1e187686368fe89c5b71b3162c094f463d0e5ec4eaca567 corroborates the 128 captured bytes beginning at segment 4 offset 869c. Every differing byte lies in an ordinary segment-selector relocation and matches its current mapped selector. In particular, the byte at 8700 is the high byte of a correctly relocated selector at 86ff, immediately before eb 29. This is code being read as a far pointer, not evidence of a missing selector relocation at the operand. This comparison is limited to the saved window: the full current generated child image was not hashed or copied.

The unresolved producer is the earlier construction or overwrite of the stack object's 005f:869c pointer. The captured LES, address calculation and far-call rejection agree with their inputs. These observations do not identify which earlier CPU instruction, API or memory behavior diverged from Windows. A future separately authorized observation should retain the writer/history of that object and its caller frame. There was no second original attempt, speculative selector substitution, guest edit, semantic fix or installation/gameplay claim.

Termination and recovery

The validated quit completed at 15:30:32.666 UTC, followed by the final readout and PAGE CLOSED +136.8s. The outer exec session returned 143. The caller's terminal/cleanup callbacks and Chrome exit observer did not write receipts; the signal source and actual probe/Chrome exit codes are unknown. This is retained explicitly, not reported as an ordinary successful browser exit. No platform cybersecurity rejection occurred.

Subsequent inspection found all three own runtime PIDs absent and both ports refusing connections. The residual owned Chrome profile was explicitly removed after verifying its recorded PID was absent. Four locks were free, with more than 9 GB free disk. No assigned boxes existed: box ID/state/expiry N/A; coordinator lifecycle ownership is unchanged. No subworkers or public deployment were used.

Recovery uses the tested Git SHA, relevant binary diffs, this new note's patch, and retained small helpers, screenshots, console output and scalar fault records. Git remains the source archive. The original failure is preserved in live-far-fault.json, live-operand.json and live-crash-detail.json; the current child pointer's producer remains the concrete diagnostic blocker.