Caesar III normal-browser city growth: authored preparation stopped by disk guard

Evidence: scratch/runs/20261118-caesar3-current-browser-occupied-housing/result.json. No Caesar III original was launched. Connected occupied housing, positive population, natural food/building consequence, Save and native Quit remain unproved. No previous occupied-city or Save claim is reused; all earlier closed scopes remain closed.

A fresh authored Win32 GDI window ran in normal production browser Workers from published commit 564c8c3bf58102262b4bdf203e63e75a16050f92. The immutable Git-backed source graph excluded foreign WIP. Its compiled module and actual executing main Worker bytecode both hash to a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8. The personally reviewed whole browser startup shows the authored blue window and input instructions. Current-display readback metadata binds the same native window, process, document and pixels. The actual 2D route has renderApi none; no GL presenter was required. The prepared route included printable C, pointer DOWN/UP, drawn menu and Escape Quit, followed only after qualification by unchanged registered caesar3_demo and a fresh Assignment 1.

The finite independent disk watchdog sampled 1,446,531,072 bytes free at 2026-11-18 09:47:18.499 UTC, below the mandatory 1.5 GB floor. It signalled the enclosing launcher. Its inherited shutdown wiring did not synchronously close the child admission: 13 ordinary CDP requests followed that disk boundary, including printable C DOWN. The independent exact C UP completed with physical state zero. A later whole screenshot visibly says “Printable C received”, but all post-guard evidence is excluded from control qualification. No clean shared UNKNOWN closure, pointer/menu/Quit qualification, or product defect is claimed. The child exited 1 at 09:47:28.659; native supervision reports no leak. There was no replay or scope reopen after discovering this closure gap.

Actual invoker, enclosing launcher, watchdog and independent release waits have both EOFs. The enclosing launcher was terminated by SIGTERM. Browser and intermediate supervisor close/EOF receipts were not produced; those custody gaps remain explicit rather than being inferred from the native receipt. Filesystem/process inspection subsequently found all 32 recorded owned PIDs absent and both owned ports without listeners, without querying an endpoint. Ephemeral browser profile and Xauthority were removed after preserving unique evidence. Only selected whole screenshots and the unique tested module remain.

Preparation hashed 38 top-level installed media files; those were rechecked unchanged. A complete recursive post-closure inventory records all 114 files, including Wavs. Whole-media preflight was not qualified and the game never requested these files. No media copies, guest/config/registry/save changes, forced polling/focus/clock/speed, synthetic city state or production edits were made. Initial/final relevant foreign binary diffs are identical. No runtime repair, focused product test pass or locked build is claimed; this is a documentation-only bounded preparation result.

The browser control harness also imported the other owner’s live tools/revolt-display-admission.cjs, which is absent from the selected production Git pin. Its exact body is retained as an added-file patch with a closure-time hash. This was a harness input, not a production graph edit; its published provenance and full control qualification are not claimed.

The concrete blockers are shared disk headroom and the run-local watchdog's launcher-to-child closure gap. A future attempt needs sufficient actual disk margin plus a fail-closed disk boundary proven on authored controls before original admission. Startup/gameplay/Quit/cleanup bounds were predeclared; none was extended. No agents, workers, boxes, Python, Wine, deployment or held title actions were used. Box ID null, state NOT_CREATED, expiry null, stop receipt null.