Conan first-store cost controls, 2026-11-17

The replacement is rejected. The single-resident specialization costs 1.915–2.007x its matching standard baseline on the generated fixed-work direct callback, exceeding the unchanged 1.20x criterion. This is a narrower route than the required full browser shell. It does not establish a low-cost observer or unblock original gameplay. No original was launched, no singular check changed, and no emulator defect or CPU/timer failure was established.

Evidence is in scratch/runs/20261117-conan-current-low-cost-first-store-gameplay/; result.json records identities, limitations, actual waits and publication. The historical denied SSH operation and archived runtime were untouched. No boxes, workers, downloads, original-media copies or frame series were created. This follows the origin rejection, cost investigation and shipped loading boundary.

Tool and measured variants

tools/conan-first-store-cost.cjs reads the five retained observer transforms without changing them. Its isolated loader redirects their store-sites.json write into the new output directory. An asserted adapter accounts for the current compiler's already-unconditional native CALL/RET seams; it changes only the inherited ENTRY-gate count assertion. No production WAT is edited. All variants explicitly report qualificationPass:false and originalLaunchPermitted:false.

The tool compiles the canonical WATX closure and protected-writer transform; compatibility modules additionally use the existing dispatch trampoline. It records module, closure, transform patch and historical helper hashes, plus the protected function-index map used for native disassembly. Commands are:

node tools/conan-first-store-cost.cjs compile baseline OUT/baseline
node tools/conan-first-store-cost.cjs compile full OUT/full
node tools/conan-first-store-cost.cjs compile single-live-resident OUT/candidate
node tools/conan-first-store-cost.cjs compile single-live-resident OUT/compat compat
node tools/conan-first-store-cost.cjs workload OUT/candidate/module.wasm OUT/timing.json 3
node tools/conan-first-store-cost.cjs verify OUT/candidate/module.wasm OUT/bits.json
node tools/conan-first-store-cost.cjs verify-lifetimes OUT/candidate/module.wasm OUT/lifetimes.json
node tools/conan-first-store-cost.cjs native OUT single-live-resident v8
node tools/conan-first-store-cost.cjs native OUT single-live-resident sm

OUT is a new run-local directory. The retained generated PE recipe remains scratch/runs/20261106-conan-zero-world-origin/native-control-fixture.cjs; there is no original launch command in this tool. Historical ignored helper inputs must be retained at their recorded paths and hashes for recovery.

The first specialization handles only epoch 1. The final specialization uses direct slot-1 globals while that remains the only allocated resident slot, including reuse with unchanged generation counters. Allocating another slot permanently restores the original generic scan. Both retain translation, overlap/span checks, invalidation, bounds, snapshots and completeness refusals. The ablations deliberately remove individual mechanisms and cannot qualify:

Direct callback variant Wall median ms CPU median ms
baseline 451.18 392.45
inherited full observer 1644.93 1470.01
omit store-hit test 602.92 447.50
omit resident scan 1286.71 484.16
omit snapshots 2083.78 1698.05
restore pending ENTRY gate 2153.29 2149.06

These sequential ablations are affected by shared host load. They locate promising mechanisms; their wall-time differences are not exclusive cost attribution. Profiling was a separate cell, excluded from timing. Its samples include linear resident getters, private register-store wrappers, translations and dispatch handlers. Correct copied matrices say nothing about CPU/timers.

Three alternating baseline/final-candidate pairs, each with one discarded warmup and three measured repetitions, gave wall ratios 2.00715, 1.91464 and 1.94677; CPU ratios were 2.05073, 1.91369 and 1.94659. Each executed 15 million native guest add/store iterations and produced exactly 45,000,000.

What the controls establish

The standard baseline, inherited full observer and final candidate share source closure 25188585292814670496007b9bbfffbb865832978c19d1bed49655c85f8c2ae3. Their recorded Git commits differ because unrelated publications happened in the shared tree; closure and module hashes, not HEAD alone, identify the tests.

Module SHA-256
standard baseline 9be4491daf35df204ca50486165952dad7bb5a9766828374083a6d16edde3a92
compatibility baseline 915c2c7ed533a7423968daca502e0bdc83ca584aa6a6655bb74cde61016505f1
inherited full standard 9e2df0fdb63e620e2b03c0bd858cb955dc4f1319682e220e30f8faa91f66d567
final candidate standard 39407f1a5be74d4661a5e738979fa8906e27453d4654705b82e03b6f0b0359dd
final candidate compatibility 15226003444ac91fdfc3719f5ab4ee8eca18199c4d4da250642a8823ba37dea3

The final compatibility candidate has closure 8cb4a168…; the compatibility baseline has 25188585…. A foreign ordinal change landed between their compiles. Their import/export contracts and tested values agree, but they are not a matching-source compatibility proof. The final locked canonical build also used the newer closure and passed; it does not retroactively qualify the older timing modules. Further foreign WORD-width changes followed that build.

check-controls.cjs in the run verifies module/helper hashes, contracts, final exported register/lazy-flag/FPU values and selected guest-data hashes. Full and specialized observers have identical complete trace-stream hashes for each tested scalar workload. Separate captures map adjacent pre/post packets to actual guest instruction PCs, with matching owner slot/epoch and exact bits:

PC Store Owner offset Before → after
431955 MOV 04 3f800000 → 00000000
431958 MOV 38 3f800000 → 00000000
43195b MOV 68 3f800000 → 00000000
4310a7 MOV 04 00000000 → 40000000
4310b0 FST 38 00000000 → 3f000000
4310b3 FSTP 68 00000000 → 3f000000

The generated two-owner control observes births 1/2, destruction of 1, reconstruction as generation 3, then destruction of 2/3. Its destructor is a generated scalar stub, not the original hot destructor implementation. Existing consumer guards still refuse absent constructor history, unknown instruction mapping, owner mismatch and selected native writers lacking full snapshots. No nativeEntries == 0 result is treated as positive coverage.

Complete architectural passivity, nesting, REP/SEH, selected hot native writers, actual destructor coverage, full-shell cost and matching-source compatibility qualification remain unproved. Those gaps independently refuse an original.

Native code review and raster clarification

Matching-module x86-64 baseline/full/final-candidate disassembly was captured with tools/wasm-native.js for V8 TurboFan and SpiderMonkey Ion. Full V8 cww_hit has a 0x48-byte local frame; specialization reduces it to 0x40 but retains translation/helper calls and the generic fallback. The full observer's linear getter comparisons and private register-store instrumentation appear in hot code, with extra frames/spills. SpiderMonkey retains calls and a 0x30-byte frame in the scan. Native files and per-function byte ranges are retained.

Static V8 is 15.7.57; measured Node embeds 13.6.233.17-node.50. SpiderMonkey is JavaScript-C158.0. This is an explicit timing-engine mismatch, not matched Node native-code proof. The separate shell execution attempts failed during PE load on an unexpected beep_process_begin import; both processes were waited and both EOFs drained. Those harness failures do not diagnose an emulator defect. There were no ARM64 measurements or captures.

The 18:27 same-task clarification suggested investigating finite collapsed solid triangles before demanding source construction. Microsoft's D3D9 rasterization rules describe triangle filling separately from point/line rules. OpenGL 4.6 sections 14.6.1–14.6.6 likewise distinguish polygon fill, point/line modes and multisampling. These contracts motivate a narrowly conditioned semantic test; they do not establish Conan's actual fill/blend/shader state.

The historical conan-shipped-loading-boundary/original/first-render-failure.json is absent. Retained validation/review/result files prove copied zero components and the image boundary but do not supply that complete draw state. It was not reconstructed from defaults. No native raster semantic control or concrete raster defect was proved, and no guard suppression was implemented.

Next work needs either a substantially cheaper design with positive native writer/destructor coverage and matching-source standard/compat full-shell controls, or recovered actual sequence-keyed draw state plus native semantic proof for the narrow raster hypothesis. Original source construction and player movement, a second visible response and native Quit all remain UNKNOWN.