log builtin silently clamps inputs below 1e-10 to ln(1e-10)

[InauguralPhysicist]

Found by phugoid rung 0 (blind-critic round 15; ledgered as phugoid GAPS.md G3).

Verified on v0.41.0:

print of (log of 1e-15)   # -23.025850929940457
print of (log of 1e-10)   # -23.025850929940457  (identical — clamped)
print of (log of 2e-10)   # -22.33270374938051   (correct)

Presumably the finite-by-construction design (no -inf), but the clamp is a SILENT plateau: phugoid's exponential-decay estimator fit log-domain samples that had decayed through the clamp range and returned a confidently wrong answer — measured 4-14% error inside a written 2% tolerance — because a relative sample floor admitted clamped samples. Any consumer doing likelihoods, entropies, or decay fits in log space hits the same plateau with no signal.

Questions for the runtime:

1. Is ln(1e-10) the intended floor, and is it documented anywhere? docs/BUILTINS.md doesn't mention it.
2. If it stays, a doc note (and possibly a lint hint for log-of-possibly-tiny) would have saved the hunt.

phugoid's local mitigation (absolute 1e-9 sample floors in both log-domain estimators + an amplitude-0.001 regression row) is in InauguralSystems/phugoid — measure.eigs and GAPS.md G3.

Co-authored during the phugoid rung-0 build loop.