Engineering, operations & design history

Fresh-browser cost-only counterpart

Full repository source: docs/art-direction/display-cache-history-20261104/clean-cost-20261104/review.md

Technical documentation includes historical decisions, proposals and verification reports. For playing instructions and practical rules, open Gameplay.

Fresh-browser cost-only counterpart

24 completed and 0 failed cases cover two counterbalanced repetitions of desktop/phone layouts at client-selected 128/192/256 MiB native-cache budgets. Every case starts a fresh owned browser. Diagnostic canvas/image reads and retained observer source-image references are absent. The existing natural town and four canonical biome sites use the same protected modules and source hashes as the separate adaptive fidelity study. No world command or production-world mutation is used.

Native MiB	CSS width	Control timed seconds, median	Private timed seconds, median	Control browser CPU seconds, median	Private browser CPU seconds, median	Control sampled browser private MiB, median peak	Private sampled browser private MiB, median peak
128	1440	87.18	84.59	89.273	86.055	3077.2	3103.1
128	390	65.29	61.66	64.688	57.484	2126.0	2116.7
192	1440	86.92	89.58	88.414	95.773	3065.7	3053.7
192	390	63.04	61.49	62.086	59.055	2111.8	2072.6
256	1440	86.05	84.70	85.430	85.953	2983.2	3016.7
256	390	65.24	66.03	64.055	65.250	2076.4	1996.8

These are descriptive paired observations from two repetitions, not statistical or whole-world performance acceptance. Wall totals sum 17 nonoverlapping measured operations per completed case; they exclude instrumentation gaps and GPU queries. OS CPU counters survive document resets but exited-process final CPU remains a lower bound. Browser private commit is summed across owned browser processes at each sample; working sets can share pages, and CPU/private/working-set/GPU values must not be added as unique physical memory. GPU checkpoints occur outside timed operations and are not peaks. The owned server is shared, and machine background effects and server warmup are not eliminated. DeviceMemory hints do not impose hardware RAM limits or certify phones.

The run records 408 metadata observations, not new pixel-fidelity frames. The separate 204-frame study remains the pixel evidence. Exact source input/transform history mismatches against that study: 0; every mismatch is retained for review before comparing equivalent work. Primary-material invocation does not isolate final material pixels or approve art/nonrepetition/contact/state.

All failed cases, costs, source inputs, ownership receipts and prior terrain/Hex failures remain preserved. Duplicate progress/current-case bodies are retired only after exact equality with the final report is proved. Original artwork, public renderer, source contracts and saves remain unchanged. Full cost/aggregate memory, relevant lifecycle/crowded-context, physical hardware and whole-artwork qualifications remain open; no approval is granted from this report alone.

Evidence: paired summary and limits validation-summary.json, all measurements world-results.json, owned-resource terminal world-owned-resources-terminal.json, retention retention.json.