experiment: retract the traversal attribution's headline, and say what it can show

[shinaoka]

The first version of src/bin/traversal.rs (merged in #98) rebuilt the case's operand layouts from the campaign row's axis record. Two m axes have the same extent, so my reconstruction swapped them and measured a different contraction — its plan folded m into two axes and ran in 11.6 ms, and the README published that as "a plan-level traversal decision is worth 3.3×". That is wrong and is retracted here.

What the corrected binary shows

Built from the harness's own la/lb/lc (read from the sized case, not reconstructed), the plan is identical to the campaign's on every recorded field — family avx512.f64.real.24x8, mr/nr 24/8, blocking 264/1536/256, StaticGrid{pm:0,pn:0}, align_c_lines: false, and the same three m axes at A-strides (92160, 48, 1) with the contracted axis at 1920. So there is no plan difference to attribute.

variant	ms	GFLOP/s
corpus-a-corpus-d	52.12	6.79
corpus-a-compact-d	51.77	6.84
compact-a-corpus-d	50.38	7.02
compact-a-compact-d	50.25	7.04

• No layout swap moves the case (all four within 4%), so the "interleaved panel" story the binary was written to test does not hold for it either.
• The binary is 1.37× slower than the campaign on the same case (52.1 vs 38.1 ms) with the same plan, dimensions and layouts, so the remaining difference is in the execution path — this binary calls execute_slices, which builds views per call, while the harness calls execute_raw on raw pointers.

Consequences

The honest state is that this binary cannot attribute the case's cost yet, and the phase-share measurement it was written for has to be made on the harness's own path. The README and the binary's header now say that, and the residual column is marked as meaningless (it compares differently permuted buffers).

This also removes the basis of the "change the planner's axis ordering, it is worth 3.3×" proposal that was made on the strength of #98. The candidates that remain measured-and-excluded are unchanged: micro-kernel, blocking model, blocking values, orientation, partition, write-back mode, output alignment, row-block shape.