[Elvand-Lie]

This Pull request:

Changes or fixes:

TProfile3D::BufferEmpty mishandles automatic axis limits.

With fixed X and Y axes and automatic Z, every entry ends up in the Z overflow:

TProfile3D p("p", "", 4, 0, 4, 5, 0, 10, 3, 0, 0); // automatic Z limits
for (double z : {101., 115., 139.})
   p.Fill(0.5, 1., z, 2.);
p.BufferEmpty();
// master: Z axis stays [0, 0], all 3 entries in the Z overflow, GetMean(3) == 0
// TH3D with the same binning and fills: Z axis [100, 140], GetMean(3) == 118.33

With automatic X, the fixed Y and Z axes are rebinned as well, so they no longer have the limits the user gave.

The outer condition only looked for automatic X or Y limits. Once inside, the block took the limits of all three axes from the buffer and let FindGoodLimitsXYZ choose new limits for every axis, fixed ones included.

The fix ports the per-axis logic of TH3::BufferEmpty:

• each axis is checked for automatic limits;
• only automatic (or extendable) axes take their range from the buffer;
• fixed axes keep their limits and pass their bin counts to FindGoodLimitsXYZ.

The buffer layout of TProfile3D, five values per entry, is kept.

Two tests in hist/hist/test/test_TProfile.cxx:

• TProfile3D.BufferEmptyFindsAutomaticZLimits: automatic Z with fixed X and Y. The Z axis contains the filled values, no entry is in a Z underflow or overflow bin, the Z mean is (101 + 115 + 139) / 3, and X and Y keep their limits.

• TProfile3D.BufferEmptyPreservesFixedYAndZLimits: automatic X with fixed Y and Z; Y and Z keep their limits.

• Both tests fail on master and pass with this change.

• ctest -R '^gtest-hist-hist-' passes: 33 tests, 0 failures, on Linux with a minimal build.

• Also checked against TH3D: axes, labels, per-bin contents and entries for automatic, fixed, labelled and extendable axis combinations, including repeated and forced BufferEmpty calls.

• Not run: the full ROOT test suite, roottest, and platforms other than Linux.

Checklist:

• tested changes locally
• updated the docs (if necessary)

Codex (gpt-6.1-sol) implemented the fix and the tests; Claude Code (Claude Opus 5.5) found the bug and validated it. I reviewed the full diff and the tests.