v2.21.0

The array-returning statistics land, which closes the last gap the README
listed. A real spilling bug turned up while building them and is fixed here too.

New

• Array-returning statistics: FREQUENCY, MODE.MULT, LINEST, LOGEST,
  TREND and GROWTH, plus TRANSPOSE. They spill across the grid on
  recalculate exactly as SEQUENCE and FILTER already do.
• LINEST and LOGEST handle several predictors, not just one. Coefficients
  come back right to left with the intercept last, matching Excel, and a fourth
  argument of TRUE adds the five-row statistics block: standard errors, r
  squared, the standard error of y, the F statistic, the degrees of freedom,
  and the regression and residual sums of squares. A third argument of FALSE
  forces the fit through the origin.
• TREND and GROWTH predict at new points, or reproduce the fitted points
  when new_x is left out. Omitting known_x uses 1, 2, 3 and so on.

Fixed

• Editing a spilling formula produced a spurious #SPILL!. The anchor's
  spill range was recorded only on its FormulaCellValue, so replacing the
  formula through updateCell, which is how you edit one, threw that record
  away. The previous spill's cells were then never cleared and blocked the new
  result, and once a cell reported #SPILL! it stayed that way. The workbook
  now remembers each anchor's range itself, so growing SEQUENCE(2) to
  SEQUENCE(4), shrinking it back, or replacing it with an ordinary formula
  all behave. This affected every dynamic array (SEQUENCE, FILTER, SORT,
  UNIQUE) since they shipped, not only the new functions.

Notes

A correction to 2.20.0. That release said these functions "need a formula to
spill a computed block rather than read one from the sheet, which the engine
does not do yet". That was wrong. SEQUENCE has always built a computed block
out of nothing and spilled it, so the machinery was already general and the
functions only ever needed writing. The README and the function reference
carried the same wrong claim and are corrected.

The regression sits on one shared least-squares fit, solved through the normal
equations with a Gauss-Jordan inverse and partial pivoting, so LINEST,
LOGEST, TREND and GROWTH cannot disagree with each other. Predictors that
repeat each other leave the system singular and report #NUM! rather than
returning an arbitrary answer. The suite checks the new functions against the
scalar ones that take a different route: LINEST against SLOPE and
INTERCEPT, TREND against FORECAST.

That takes the engine to 321 function names.