docs: the smooth-union licence measured at exactly +k margin, and the refusal kept anyway (OQ-54)
docs: what a blend costs - the culling licence priced at 443x, and the demo printing its own refusal (ADR-024, OQ-54)
docs: the descent is worth 443x where one bound was worth nothing, and compaction's lever turned out to be capacity (ADR-024, OQ-49, OQ-53)
docs: the evidence manifest - measured numbers in one place, checked by the docs gate and by CI
docs: the cull has a caller - the boot marches twice, asserts the frames agree, and reports its own limit (R1, ADR-024)
docs: assetc watch - atomic replacement, and a broken edit that never looks like a success (AS1, OQ-48)
docs: AS2 becomes a cross-machine gate, and the two harness bugs that made it red on three machines that agreed
docs: a frame the caller owns, and what the ban on unsafe lets the allocation audit actually measure (AS4, S4)
docs: JavaScriptCore agrees - Safari on the ARM64 runner produces the native digest (P1.5, S3, OQ-41)
docs: paragraph break before the frame-time caveat
docs: the build-time union normalisation, its equivalence tests, and where the sort must stop (AS1-AS5, ADR-024)
docs: tree order is worth 2.6x-3.0x of the culling, measured on one scattered scene emitted twice (ADR-024)
docs: keep the culling baseline's column name consistent after the rename (ADR-024)
docs: the node-visit tally and the first measured number for the per-node skips (R1, R2, ADR-024)
docs: the per-node skips, the smin refusal they hit, and the metric that cannot see them (R1, R2, ADR-024)
docs: fix the code identifiers a shell substitution ate out of the culling section (ADR-024)
docs: the enclosure cull, its measured saturation, and the 3x estimate it replaced (R1, R2, ADR-024)
docs: date the boot's line-count claim to its era, since the count moved twice after it was written (P1.5)
docs: the browser job is in CI on three architectures, and the run that taught the shim to derive its imports (P1.5, OQ-41)
docs: the boot's roots now print inside a browser, and what that does and does not close (P1.5, OQ-41)
docs: mirror gate P1.4's directory, its measured cost, and OQ-52 (P1.4, OQ-49, OQ-52)
wiki: spike S7's first run - latency is distance over ring, cost is a curve not a cliff The three-stage model's unvalidated half is now measured on a 512-region integer river: the mouth dries at commit 501 with ring 1 and 64 with ring 8, and the work against simulating the whole river rises 57% -> 95% as the ring widens. So 'without loading region B' holds, 'within one epoch' does not by design, and the warning about cost exploding does not survive the column - a wider ring pulls the same work forward. Recorded with the two mistakes the run caught, because they are the useful part: a field whose lower river was already dry at equilibrium (so four distances could never be seen to dry), and a commentary written before the table printed that claimed the work barely moves. OQ-01 narrows from 'can causality propagate' to 'how much multi-epoch lag will we accept', and SM1's two-clients half still waits on A-04 and A-05.
wiki: the WASM parity is a three-architecture matrix, and arm64 V8 matches its own native build Run 38270620391: `parity ok on arm64` from macOS, `parity ok on x86_64` from Linux and Windows, 13 lines identical each. Stated as what it is - the instruction set in the gate's M1-browser row, executed by V8 rather than a browser - because a green tick here is otherwise easy to read as four of four. Also repaired a sentence the previous edit had broken mid-clause ("...every CI architecture and, runs under..."): a wiki page with a dangling conjunction is the same class of defect as a dead link.
wiki: the roots run inside a WASM engine, and the blocker we cited for it was a misreading docs/11's claim was that WASM execution needs the export-lint decision because a runner would have to be wit-bindgen output. Both premises are true and the conclusion is not: a WASI command exports _start from the Rust runtime, so nothing in this repository writes an export or an unsafe. The engine now builds for wasm32-wasip1, runs in CI under Node's WASI, and diffs identical against the native build - 13 lines, EPOCH_0 91b97a52... among them, main green in 38265695965. The step fails on empty output, because an empty diff between two empty files is a green check that means nothing. Home, Open-questions and the status log all repeated the old claim, so all three are corrected rather than one of them quietly updated. Stated alongside it: this is a fourth kind of target, not spike S3's fourth device - M1-browser still means a browser on an M1 - and the marched f32 values riding in the same lines are presentation, not evidence (ADR-017). OQ-41 stays open for mods exporting functions; it was never what blocked this.
wiki: precise about which column varied - 0.43-0.78 ns per node-step at 31 nodes against 1.58-1.62 at one The previous wording quoted 0.41, which is a marginal-column value rather than an ns-per-node-step one, and range-mixing the two columns is how a measured page turns into a vague one. Same conclusion, stated against the right column.
wiki: three runners print 659,801 and 627,890 march steps identically - so the budget binds node-steps The node-cost sweep runs on every gate job now, and the two halves of the result separated cleanly: the step counts repeat digit-for-digit on x86-64 Linux, x86-64 Windows and ARM64 macOS, because a march step belongs to the field along the ray rather than to the tree or the hardware, while the nanoseconds moved as nanoseconds do - ns per node-step between 0.41 and 1.62 depending on the runner, for identical work. Which is why docs/04 states the rule as bind-node-steps-not-milliseconds: a budget quoted in nanoseconds would be a claim about one machine in a project that promises several.
wiki: the frame law is steps x nodes, and OQ-50 blocks the visible edit raster_probe's tree-growth sweep: the march step count barely moves with tree size (659,801 steps for one node, 627,890 for 31) while the work per step scales with the node count - about 1.5 ns per extra node per step here. Recorded with the caveat that the timings varied 2x between runs of the same binary while the counts repeated exactly, so a budget belongs in node-steps rather than milliseconds. The second half is a gap, not progress: projecting an edited voxel into an A-07 cube needs to know which world span GlobalVoxelPos names, and no doc says. Nothing hashed depends on it, so it is not a determinism question, but the raster path, doc 07's collision and the future base terrain must agree - ADR-021's single coordinate authority. Filed as OQ-50 on the Open-questions page rather than guessed in code, and gate P1.7's row now names it as why the frame-visibility half has no implementation.
wiki: spike S2 part three - where P1.3's line falls, and what batching across regions is worth Roadmap gains the sweep table and the two findings that contradict what the docs asserted: spreading 65,536 edits over eight regions computes 548,192 digests against 483,285 in one (13.4% MORE work, so that OQ-49 lever is parallelism rather than reduction), and 1 to 16 slot rows commit in 4.10-4.43 ms, so the live -region count is not the cost driver. The 1 ms line sits between 128 and 512 edits per region per epoch. Open-questions' OQ-49 entry was two spikes stale - still quoting 4.9 ms for the pass and 93 ms for the commit, from before the O(N^2) resolution pre-pass was removed. It now carries the current range, the packet command, and the caution that CI repeats the digest counts on three runners while its timings disagree about which arrangement is faster. Which lever moves remains the owner's call.
wiki: the boot marches its own grid, and three architectures print the identical line (a2465dd, run 38250242415) 590 of 3,072 rays find surface; 243,008 evaluations charged, 98.1% of them spent by rays that found nothing. The counts are integers because the march and the normal use only add, multiply and sqrt, so they are comparable across machines - and x86-64 Linux, x86-64 Windows and ARM64 macOS do agree byte for byte. Recorded as a diagnostic next to the number, not as gate evidence: a pixel buffer is not hashed, and ADR-017 keeps frame hashes informational.
wiki: the field becomes a picture on a CPU, with the first measured frame-cost shape Mirrors 237c916. render::raster marches a pinhole grid and reports 45.5 evaluations per pixel, flat across three resolutions on one tree, with the 33.0% of rays that find nothing charged 69.6% of all evaluations - ADR-024's argument as a measurement instead of a derivation. The page states plainly that this is not R1: no GPU, no WGSL, no BVH, and a normal-mapped colour that is not a material model because OQ-10 has not settled lighting. Counts moved to 118 in twenty suites, run 38249012218, seven jobs green on all three runners.