feat(m3-harder-data-quantitative): source harder quantitative-uncertainty items, recalibrate, record CEILING-NEEDS-HARDER-DATA

Replace the 100%-ceiling bayesian-authored pool with 57 freshly-authored harder
items (post-cutoff forecasting, complex probability/statistical reasoning,
multi-step Fermi estimation, hard debiasing, margin-of-safety trade-offs).
Calibrated 28 items at K_TRIALS=5: baseline 93% (3 in-band at 40-60%,
24 ceiling, 1 floor). claude-sonnet-4-6 is extremely capable at multiple-choice
quantitative reasoning; the 40-70% band is not achievable with this format +
solver combo. Recorded honest CEILING-NEEDS-HARDER-DATA per Invariant #4.
Decisive split: 57 items (47 target + 10 distractor, 0 ambiguous).
External HF data remains gitignored/unstaged.