Guarded printable USER input

Evidence: scratch/runs/20261117-caesar3-native-printable-text-city-progress/result.json.

The predecessor's original run confirmed physical C DOWN but left Caesar III's name field empty. TranslateMessage deliberately relies on the browser's printable character event. Physical DOWN is not a character receipt.

The CLI now accepts keytext:VK:CHAR_CODE as one finite guarded intent. It calls the existing physical DOWN handler, then the existing browser character handler. It never derives a character from a virtual key, changes translation, sets guest text, or posts a guest command. Independent release still owns physical UP.

Both produced queue objects must match the exact DOWN/CHAR pair, including the shared keystroke lParam. Each needs its own opaque token and an actual matching owner USER API return, in order. The receipt remains pending after DOWN and reports both native messages only after CHAR. Exact owner PM_NOREMOVE returns remain valid; a subsequent removal cannot confirm the pair twice. Printable proofs additionally require the bound main HWND, preventing a sibling HWND on the same actor from claiming a character queued with HWND zero.

The initial qualified scope is explicit printable ASCII on a plain main guest window, with no native modal, native edit, modifiers, menu interception, capture, or other held key. Unsupported characters and native control routes refuse before effects. Full existing focus, generation, actor, memory, queue, device, invocation and immutable deadline checks remain in force. The printable handler identity is pinned alongside the physical handlers in the owner snapshot.

Authored real x86 fixtures publish DCU after UP, one marker per native DOWN/CHAR/UP callback. Publishing on UP deliberately preserves the strict pinned scene while CHAR is pending; changing a caption on DOWN correctly refuses the later character. Cooperative and CPUWorker CLI runs check exact callback order, no translation duplicate, distinct native tokens, matching lParam, independent release and unchanged 500/1750/3000 ms bounds. Separate real main and CPUWorker owner execution exercises the production native broker and mismatched actual character proofs. Guard tests cover cancellation, lateness, actor/thread/memory, focus/modal/reuse, owner retirement, missing characters and reversed order. The exact authored controller also reaches the public DCU consequence.

Original city, assignment consequence and native Quit require separate visible evidence from the single authorized fresh invocation. The predecessor's empty name field, failed Escape exit, missing native wait/EOF and unproved old timeout cause are preserved as facts; this change does not reinterpret their filenames or revive any old UNKNOWN endpoint.

The subsequent outer adapter reuses the GL worker's published ParentProcessSupervisor at the actual command boundary (exec python3 ... supervised). It pins that supervisor, the underlying Boids helper and the Caesar helper to a chosen current published revision. Preparation remains 240 s; the runtime and independent runtime watchdog remain 2400/2420 s. The prepared handshake follows compilation, full authored qualification and (for an original) all 114 recursive media hashes. Inner native timing and input limits are unchanged. Later unrelated main publication may advance the branch without refreshing the sealed source/module/helpers. Actual parent-process loss still causes shutdown. This integration does not assign a historical signal cause.

The single fresh original at published 2639be4c01f6a7483fe3575d55fb430a3cdd9769 passed exact source/module/helper/main/Worker/OS-child qualification and all 114 recursive media hashes before the runtime clocks. Caesar visibly accepted exactly C: the DOWN and CHAR had distinct native tokens, matching HWND/thread/lParam, and actual owner Peek PM_NOREMOVE returns. Confirmation took 88.867 ms, independent UP 245.942 ms, and complete immutable observation 375.150 ms. Personally inspected complete images then showed Assignment 1 — Aventine and the actual city landscape (Dn 2000, Pop 0, Jan 500 BC).

The next housing selection failed the five-second gameplay image freshness preflight, before its intent or physical DOWN was created. The controller permanently closed the endpoint as UNKNOWN. No later endpoint request, rebind, replay or second original was made. Housing, road, food consequences and native Quit are not proved. Host cleanup is distinct: actual native wait returned zero with both EOF and hostCleanup: true; outer actual wait and both independent watchdogs returned zero with both EOF and no remaining owned processes. A zero cleanup status is not a native Quit receipt. No box was created.