Legacy mip continuation: MIPLINEAR

Base: fetched main ee3a71e3398a98a360d02c25d02bfeb544f0e14f, including combined correction fbf684a1796e47a0aa5c1552bb38dd743cd3c16a. The first demonstrated remaining defect is MIPLINEAR (0x08). Only fill_primcaps.dwTextureFilterCaps changes: 0xFB to 0xF3. DX7 builders inherit this as 0x030003F3. NEAREST/LINEAR, DX7 MINF/MAGF POINT/LINEAR, the prior cube/address corrections, and all other caps remain.

The Microsoft D3DPRIMCAPS contract requires mip-level selection and within-level bilinear sampling for MIPLINEAR. The saved SDK headers in the evidence identify its cap as 0x08 and legacy D3DFILTER_MIPLINEAR as 4. Stage filter enums differ: value 4 is not used as a MIPLINEAR request in the stage-state fixtures.

Implementation evidence is stronger than those names: the filter-selection block inside d3dim_draw_tl_triangle_textured sets linear only for value 2. The exported d3dim_gpu_describe independently writes the same boolean. d3dim_texture_sample_keyed takes one bound image and no chain, derivatives or LOD; its zero-linear branch takes one point tap. lib/d3dim-gpu.js turns that boolean into point min/mag and sets mip: 0. The legacy MIN-only request (offset 328, stage filters and legacy MAG zero) therefore fails even the within-level bilinear requirement. No claim about automatic LOD is needed to establish this defect.

node test/test-legacy-mip-support.js executes extracted production WAT and the actual JS sampler assignment:

• 36 state cases: legal point/linear stage controls and legacy point/linear/ MIPNEAREST/MIPLINEAR controls, each with stage MIPFILTER 0/1/2.
• Two explicitly bound images (4x4 black/white and 2x2 red/blue checkerboards), instrumented fixture fetches, and actual production addressing/bilinear helpers. LINEAR blends four taps to gray/purple; MIPLINEAR takes one black/red point tap. These are distinguishable image fixtures, not a simulated automatically selected mip chain.
• 90 filter fields across HAL/HEL enumeration builders and GetCaps 1/2/3/7, both primitive kinds, aligned/unaligned storage, descriptor sizes and guards. Enumeration dispatch-to-builder wiring is checked; callbacks are not executed.
• Seven mutation negatives: re-advertised MIPLINEAR, restored MIPNEAREST, missing point, missing linear, unrelated filter clearing, changed GPU mip lowering, and changed software filter selection.
• The unchanged fetched-main baseline passes the backend observations but fails specifically because it advertises MIPLINEAR.

This is extracted-source qualification with fixture state/fog/fetch boundaries. No full draw, native COM dispatch/callback, actual GPU render, game, canonical build or leased runtime was executed. The previous filter task's actual receipt, contract log, test and source were read before extending this fixture.

Scope stops here. Remaining legacy 0x10/0x20 claims are not qualified by this task; reserved bits are untouched. Modern mip-stage bits were already absent. No other GL/D3D capability audit was performed.

Evidence and standalone replay: scratch/runs/20261020-legacy-mip-support/README.txt. The independent reconciliation of commit 795d74657 is on task/legacy-mip-reconcile-oct20, based on fetched main ee3a71e33. All four existing exact-mask suites are now reconciled in tracked files: test-legacy-cube-caps-contract.js, test-legacy-texture-remaining-caps.js, test-legacy-texture-filter-caps.js, and test-d3d-device-enumeration-lifetime.js. Legacy expectations are F3 and DX7 expectations are 030003F3. Cube/address masks and point/linear behavior remain unchanged. The earlier filter suite's mutation anchor now matches F3; its MIPNEAREST and missing-point/linear mutants change one bit each (F7/F2/F1). The MIPLINEAR suite retains FB as an intentional negative control.

Combined source replay passes 115 cube checks, 90 address fields, 90 earlier filter fields, and 90 MIPLINEAR fields, plus 72 address and 36+36 filter cases and 23 mutation controls. The three stale source suites fail on the reviewed production source before expectation correction. All four corrected source suites reject the unchanged main baseline; the MIPLINEAR baseline first passes backend observations and then rejects the advertised bit.

The native enumeration suite was syntax-checked only. Native COM GetCaps and enumeration callbacks (including nested/lifetime behavior), canonical full build, actual software/GPU draws and automatic mip-level selection remain unqualified. No game result is claimed. Remaining 0x10/0x20 bits were not audited. The production diff is still exactly the original FB-to-F3 line. Main integration remains coordinator-owned.

Self-contained reviewed-source snapshot, baseline, logs, patch, hashes and JavaScript replay: scratch/runs/20261020-legacy-mip-reconcile/README.txt.