Steep software sampler stages

Evidence destination: scratch/runs/20261106-legacy-sampler-stage/.

This investigation follows the rejected diagnostic. The retained public sample is integer pixel (30,1), with independent RGB [103.220778044,113.924726814,97.215804982] and tolerance 2. Geometry, the oracle and tolerance stay unchanged.

Diagnostic declarations are prepared in the included core VFS fragment under all three normal aliases. The root remains byte-identical and include-only. The original malformed-root transform is separately replayed with a compiler entry counter before any diagnostic compilation. Runtime detection, protected exports and strict declarations remain enabled through compileClosure.

Prepared trace events distinguish lookup UV from quotient-plane UV/Q, rho squared and LOD, selected mip dimensions, bilinear taps/fractions, quantized weights, actual horizontal rows, packed bilinear colors and the mip blend. Diagnostic modules cannot qualify the shipped-module matrix.

The original malformed-root negative passed: the exact predecessor transform raised protected runtime requires an include-only root, with zero compiler entries. The corrected diagnostic compiled normally. Its bootstrap and browser both armed an exported writer, observed its rejection, disarmed it, and verified the mapped bytes remained unchanged. The focused protected closure contract also passed. No protection/compiler policy was changed.

Actual retained stages

Unchanged production again passed the original twelve samples, then returned [103,113,95] at (30,1). The diagnostic returned the same pixel and captured:

Stage Actual value
Lookup UV .25238463282585144, .23146159946918488
Quotient-plane UV .25238463282585144, .2314615696668625
Q .03249996900558472
rho squared / LOD 4.5312957763671875 / 1.0899618864059448
Mip fraction .08996188640594482
Selected dimensions 8x8, then 4x4
Level 1 fractions / integer weights .5190770626068115, .351692795753479 / 132,90
Level 2 fractions / integer weights .5095385313034058, .4258463978767395 / 130,109
Level 1 horizontal RGB rows [101,157,96], [101,37,93]
Level 2 horizontal RGB rows [126,54,114], [126,174,113]
Packed bilinear results [101,114,94], [126,105,113]
Packed mip blend [103,113,95]

stage-observations.json retains all actual taps and events. Its arithmetic checks reproduce every observed horizontal row and packed bilinear result. The first filtering discrepancy is fractional-weight quantization, followed by horizontal and vertical truncations. Ideal bilinear results using the actual traced inputs are [100.863369,114.796865,95.543219] and [127.068316,105.101568,114.135805]; their blend is [103.220815,113.924657,97.215843]. The actual packed levels instead blend to [103.249047,113.190343,95.709276], which the final conversion truncates. Thus the observed filtering stages explain the failing blue channel. The tiny lookup/plane UV difference and LOD approximation do not explain this pixel's error. This is stage evidence, not a general legacy precision bound or a claim that all perspective interpolation cases are correct.

Candidate and first new failure

candidate-repair-and-regression.patch preserves an unqualified candidate: fractional SIMD bilinear rows with one final rounded packed conversion, the reversed-gradient/steep public regressions and their analytic references, and the predecessor's GL numerator/Q reconstruction without changing raster positions. Both modules passed the locked full build. Shader, analytic integer-center, ValidateDevice, protected-closure and public-plan checks passed.

Validation stopped at an operator invocation failure at 2026-11-06T12:47:42.900099Z: I named nonexistent test/test-d3d9-fixed.js. Concurrent mirror and Worker-plan commands also omitted their required --plan flag and rejected missing runtime configuration. These errors are preserved in the focused command records/logs; they are not emulator defects. There was no retry or candidate rendering afterward.

All four owned production/test edits were restored after saving the patch and candidate modules. No repair or regression is landed. Candidate public software pixels, fixed/top-left and mirror/Worker plan checks, native sampler regressions, and every candidate solid/spatial matrix cell remain unexecuted. The next worker must first use the actual existing fixed test filename (test/test-d3d9-fixed-top-left-bias.js) and explicit plan/runtime arguments, then qualify the retained candidate's software regression and full matrix.

Identity and cleanup

Baseline, diagnostic and candidate source records pin Git f22af401717fb054b429a954b393bd4e306b5ae5; each has a Git binary diff and 4,061 source hashes. The diagnostic additionally has trace-transform.patch. No source tree or media was archived. The existing calc fixture was referenced at SHA-256 cb71fcc9726f756421519a215c39cf89efd930c9ae0fa5bbb59b64050faef52b.

Module Standard SHA-256 Compatibility SHA-256
Unchanged production 220beaba026d0aabaf25b4193a345ca6d704be64b0812a72d68adb0700ad80f2 2b180ff8bbf552a10d97ac21d8ed5bf820e712e5c658bb29f0006e623a867326
Diagnostic 24a70494fe4005dc979b6404843709c22056ce824c1ffbd64c9473a7239929d8 unchanged, bootstrap only
Candidate, build only a26cbca6cb7e1a5aa106048138248f18b610a1eed3222f526c79cee74327c962 99e657781f4cab4e2bf3c3f985f53dddd699f4e1c8082be3f1343ad0c312f785

Ordered Legacy1200, Disciples1216 and KKnD1221 FINAL releases and independent cleanup passed before the sole window, 12:44:17.103448Z–13:29:17.103448Z. Both synthetic browsers closed normally; their failure image was reviewed as a synthetic triangle only. Final cleanup verified 131 absent recorded PIDs, six refused ports, two absent profiles, four free locks and over 4.4 GB free disk, within the window. The board watcher was already absent when cleanup attempted to close it; this is recorded separately from test failures. Baseline module bytes were restored to build/; ignored build/combined.wat was generated by the candidate build and requires the next normal rebuild. No boxes were assigned: ID/state/expiry are not applicable. Runtime ownership returns to the coordinator. Evidence is published in result.json last.