Steep perspective software filtering

Evidence destination: scratch/runs/20261106-legacy-steep-software/. The unchanged production baseline reproduces the authored failure. The next diagnostic failed before producing a module, so actual internal sampler stages remain unobserved. No emulator repair or GPU candidate was applied or landed.

The authored public regression and independent oracle are retained from the predecessor. Pixel (30,1) has expected RGB [103.220778044,113.924726814,97.215804982], tolerance 2. Fresh public x86 COM drawing returned [103,113,95] at that integer center. The original twelve spatial observations passed first. Both shipped modules passed resource/bootstrap checks; compatibility rendering was not reached.

The locked full build passed at Git ee22703a6613f64690a2d1328b82ec42db98cf3e, with no tracked production/test diff. The public regression was added in memory by public-probe.cjs; baseline-harness.patch records the complete test adaptation. It retains the existing independent oracle and tolerance, and limits execution to standard software spatial drawing. The four relevant source/test hashes were unchanged since dispatch. baseline-source.json records 4,061 source pins and baseline-source.patch records the empty tracked diff.

Shipped module SHA-256
standard 220beaba026d0aabaf25b4193a345ca6d704be64b0812a72d68adb0700ad80f2
compatibility 2b180ff8bbf552a10d97ac21d8ed5bf820e712e5c658bb29f0006e623a867326

These match the predecessor's retained modules exactly. The existing test/binaries/calc.exe was referenced, not copied; its SHA-256 remains cb71fcc9726f756421519a215c39cf89efd930c9ae0fa5bbb59b64050faef52b.

Microsoft documents bilinear filtering as a weighted average with texel centers at (i+.5)/dimension, including an example whose 127.5 component is represented as 128. The pages consulted do not specify a general bound for intermediate fixed-point quantization or UNORM rounding in legacy D3D. The example alone must not be promoted into a complete precision contract.

References: bilinear filtering, linear filtering, texture filtering. The D3D11 functional specification is a different API generation and cannot establish a D3D7/9 intermediate-precision requirement by itself.

The intended stage diagnostic would record sampler inputs, quantized weights, horizontal rows, bilinear results, mip fraction and final packed blend at the failing pixel. My prepare-trace.cjs incorrectly prepended its trace import and global to the root manifest. compileClosure rejected this at tools/protected-writer-entries.js:103 with protected runtime requires an include-only root. The guard requires every root form to be an include. This is a diagnostic preparation defect, not an emulator failure. compile-trace.log and trace-transform.patch preserve it. No diagnostic module was produced, no trace browser ran, and no retry followed this first new failure.

The concrete next step is to keep main.watx include-only and put diagnostic declarations in an already included VFS fragment, preserving the protected writer pass and all three VFS aliases. That proposed correction is untested. Then capture the actual stages before deciding whether an emulator change is justified. The predecessor's intermediate-truncation explanation remains offline arithmetic; this run does not turn it into runtime causal proof. The unresolved semantic boundary remains whether legacy filtering permits packed eight-bit conversion between the horizontal, vertical and mip blends, and what intermediate precision/error and rounding rules apply. An ideal oracle mismatch alone does not settle that boundary.

All ordered FINAL releases and independent cleanup passed before the local window, 12:20:22.899604Z–13:05:22.899604Z. The only browser ran 12:20:58.001Z–12:20:59.066Z. Initial cleanup at 12:21:29.996Z verified 58 absent owned PIDs, three refused ports, one absent profile, four free locks and more than 4.57 GB free disk. The reviewed screenshot is a synthetic triangle. No boxes were assigned; box ID/state/expiry are not applicable. Runtime ownership returns to the coordinator.

Unexecuted: internal stage capture, focused repair regression, all solid matrix cells, completed spatial matrix cells, compatibility spatial rendering, direct and owning-Worker WebGL1/2 spatial rendering, and N/Q GPU candidate qualification. Original GL coverage, top-left ownership and texture lifetime were untouched. No failing test or speculative production change is committed.