A local mind, with room to choose

Implementation and plan: docs/local-first-refinement.md. Catalog model-library-1f1413fe06a343dd has six vision options and one recommended starting model. Qwen3.5-2B downloaded 1,949,063,104 verified bytes; two isolated actual image calls on llama.cpp 0.6.0/build11429 returned JSON descriptions (13.37s cold, 6.65s warm). Input was historical approved Unity capture 20261014T210245147170-f615b1986369, resized to 768x432; these are compatibility probes, not gameplay performance acceptance. Probe source: private probe-results.json and isolated ledger, 15 October 2026 Zurich. Automated scope: 322 tests passed in 16.88s on 15 October, source pytest/JUnit report, including catalog selection, persistence, privacy and budget suites. Browser fixture source model-settings-1639e5b44adb3e39 before / model-settings-0b3694fa906b49d2 after: six cards at both 1100px and 390px; recommendation count 2→1; Qwen4/GemmaE2B button routes checked. Both iterations preserved with four individually inspected approved captures. Browser fixtures contain synthetic state only. Native one-line recommendation edit is applied, with no compilation failure or current console errors reported at 22:19:07 UTC; native capture, live backend restart and actual resident assignment remain pending while the shared native player is under character-picker acceptance. Optional models are not inference-tested. Offline installer, continuous local quota policy and richer daily-life/gardening/visual work remain planned. Active game chat received documentation-standard/overview-site instructions and handed off model ownership in its commentary; idle website chat was not awakened.