Compare the signs of the two rho z scores

[arpitjain099]

Closes #279.

The rho rule in flag_ww_outliers() reads

sign(.data$z_score_rho != sign(.data$z_score_rho_t_plus_1))

The close parenthesis is one term too far to the right. Instead of comparing the two signs, it compares a z score against the sign of the next one and then takes the sign of that logical. sign(TRUE) is 1, so the term is true whenever z_score_rho is anything other than exactly 1 or -1, which is almost always.

That makes the sign test a no-op in practice, and the rule reduces to "two consecutive steep moves", regardless of direction. A sustained rise or fall then matches, and gets flagged as an outlier, which indicate_ww_exclusions() can go on to exclude from the model. Only the spike case, a steep move followed by a steep move the other way, was meant to match.

Running the same data through both versions, with log_conc_threshold set high so only the rho rule can fire:

sustained rise (same sign)       rho_threshold=0.8  current=1  fixed=0
sustained rise (same sign)       rho_threshold=1.0  current=1  fixed=0
sustained rise (same sign)       rho_threshold=1.2  current=1  fixed=0
spike up then down (opposite)    rho_threshold=0.8  current=1  fixed=1
spike up then down (opposite)    rho_threshold=1.0  current=1  fixed=1
spike up then down (opposite)    rho_threshold=1.2  current=1  fixed=1
sustained fall (same sign)       rho_threshold=0.8  current=1  fixed=0
sustained fall (same sign)       rho_threshold=1.0  current=1  fixed=0
sustained fall (same sign)       rho_threshold=1.2  current=1  fixed=0

So the change removes the false positives and keeps the true one.

Two tests added, at three thresholds each. One asserts a monotone rise and a monotone fall produce no outlier, and the other asserts the spike still produces exactly one, so the fix cannot pass by simply flagging less. The first fails on main with six failures; the second passes on both, which is the point of including it.

All the pre-existing tests in test_flag_as_ww_outliers.R pass against both versions, so nothing that was flagged before and should still be flagged has changed. I ran them by sourcing R/preprocessing.R directly, since the suite's setup.R loads the compiled package and I do not have the Stan toolchain here. CI covers the rest.