b0e9db6e

By: ValentinHirschi <valentin.hirschi@gmail.com>

W0.3 harness note: fix round (provisional factor, budget, placement scope)

Independent verification raised 19 problems; the note now:
- labels the per-CCX factor 1.14-1.17 [E, loaded, provisional], quotes the
  same-session contrast and brackets projections by 1.42-3.75 until a quiet
  session measures g1/g12/ccd12-C with >= 2 repeats; withdraws "quiet value
  expected lower" (load lowered the contended arm and the clock);
- withdraws the prepare-subtracted fill columns (raw tables instead) and the
  "DRAM flat K=8->K=24" statement (+49% with a prepare estimate removed);
- adds a quiet K=96 cycle budget (far DRAM <= 9%, uncontended transfers ~6%,
  >= 84% residual needing queued transfers) and marks the mechanism summary
  as qualitative;
- limits the placement conclusion to owner-program data under the counting
  sink (single loaded control, kernel.numa_balancing=1, no numa_maps) and
  keeps layer/index placement open for W3.5;
- leaves the context-count vs variable-map-count split unresolved;
- states GiB vs GB, the 3.5% K=1 noise floor, the workload-mix limits and
  the W1.2 differential/RSS-growth requirements.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>