fix(report): stop "not distinguishable" reading as a pass

[TMHSDigital]

The inconclusive verdict read as a pass.

ECE 0.0740 ... not distinguishable from a perfectly calibrated model at this
sample size.

That is exactly what the arithmetic establishes, and close to the opposite of
what it means. A reader skimming sees their model compared against a perfect
one, no difference found, and takes it as a clean result. It actually means the
dataset is too small to resolve the question either way: a well calibrated
model and a badly calibrated one both land inside the band on too few rows, and
the figure does not say which you have.

This is the tool's own headline verdict doing the exact thing the tool exists
to prevent, which is why it was worth its own pull request.

After

ECE 0.0740 ... INCONCLUSIVE at this sample size. A perfectly calibrated
model would often score this badly on this many rows, so this dataset cannot
tell the two apart. This is not a clean bill of health: nothing was
established either way. Collect more rows to make the question answerable.

Fixed in the report output first, not just the README, because the report
is what users read. Both call sites changed: the calibration verdict and the
baseline figure statement. The report's "How to read this" preamble now defines
the term instead of assuming it.

Tests pin the valence, not the wording

• test_an_inconclusive_statement_does_not_read_as_a_pass asserts the
  statement says "not a result in either direction" and contains none of
  passes, acceptable, looks fine, no issue, is calibrated.
• test_the_preamble_says_inconclusive_is_not_a_pass asserts the report
  carries the definition.

Four existing tests asserted the old string and were updated to assert the new
meaning rather than the new string.

Knock-on

docs/example-report.md regenerated so the committed page shows the new
wording, with its hand-written header rewritten to match. The README's four
block quotes were verbatim and are verbatim again, verified
whitespace-normalised against the regenerated report. METHODOLOGY and PLAN no
longer use the old phrasing anywhere.

Also corrects "every figure is read against its own null" to "every
inferential figure" — cost and latency are measurements, not inferences,
and correctly have no null. That was a literal overstatement.