[davidhbernstein]

Closes gap A60.

The defect

sfm() sets $sigma_u_at_bound when the fitted one-sided scale reaches the
zero boundary. Two cases reach it, and they read in opposite directions:

residual m3	meaning	correct reading
>= 0 (wrong skew)	Type I failure of Olson, Schmidt and Waldman (1980)	the boundary is the MLE; no evidence of inefficiency
< 0 (correct skew)	the data say inefficiency is present	the boundary should not be the answer; the fit is suspect

Only the first was handled.

The warning was guarded on the conjunction isTRUE(ws$wrong_skew) && isTRUE(ws$at_bound), so the second case was flagged and silent. That guard
is correct for the message it protects -- .warn_wrong_skew_boundary() explains
the Olson-Schmidt-Waldman case, which needs m3 >= 0 -- but it left the
opposite case with no message at all.

print() and summary() were worse. .sfa_report_boundary() dropped the
clause naming the skew but kept the sentence after it:

NOTE: sigma_u is on the zero boundary.
  Under wrong skewness this is the correct MLE, not a failure (Waldman 1982);
  see ?skewness_test for a p-value.

So the suspect fit was shown the sound fit's exculpation. That matters more than
the missing warning, and the comment above that very function says why: a
warning "does not survive saveRDS(), a fresh session, or a loop that
suppressed warnings"; the printed table is "what people look at afterwards".

Measured

sfm(model_name = "NW"), y_pcs_w, N = 150, seeds 1-8. Three fits land here,
all profiled with the package's own $objective (keep_objective = TRUE), not a
transcription -- NW is simulated ML, so a transcription would use different
Halton draws and the comparison would be meaningless.

seed	conv	SEs	residual m3	gain from moving sigma_v alone	$sigma_u_at_bound	$wrong_skew	warning
1	52	0/6	-1585.5	0.4771	TRUE	FALSE	none
5	52	0/6	-469.3	0.0452	TRUE	FALSE	none
7	52	0/6	-4007.9	0.0000	TRUE	FALSE	none

Two of the three are not even stationary: holding every other parameter at
its fitted value and moving sigma_v alone improves the log-likelihood. On seed
1 a point 6.5743 better exists at the same sigma_u (sigma_v
3.1328 -> 2.8300, k 0.1344 -> 0.1314), verified by evaluating $objective at
the full candidate vector. Seed 7 is stationary in that coordinate and has not
been shown non-stationary in another, so the non-stationarity claim is 2 of 3.

So a user saw a boundary flag, no standard errors, and a printed note calling it
the correct MLE -- on a fit that returned ABNORMAL_TERMINATION_IN_LNSRCH and
was not a maximum.

The fix

Both reporting paths now branch on the skew. The suspect case gets its own
message, which names the residual moment, reports a non-zero convergence code
when there is one, says the fit may not be a maximum, and states explicitly that
the wrong-skew reading does not apply. The sound case keeps its message
unchanged, verbatim.

The guard is model-generic, not NW-specific -- it fires for any sfm() model
reporting sigu or lambda -- so this covers the class, not just the three NW
fits that exposed it.

One exclusion, which is the subtle part

tHN is excluded and keeps its own reading. It already warns at
sfm.R:1612 that heavy-tailed noise absorbing the entire one-sided component is
a known property of the model rather than a numerical failure, and its row names
(sigv|sigu|nu) put it through this same generic block. Without the exclusion
one tHN fit would print both "this fit may not be a maximum" and "a known
property of tHN, not necessarily a numerical failure".

The print path excludes it too, and substitutes tHN's own reading -- including
when only the generic flag fired, because the two thresholds differ
(tHN: sig_u/sqrt(sig_u^2+sig_v^2) < 1e-3; generic: sigu/sd(resid) < 1e-2),
so the generic one can fire alone. That case previously printed the Waldman
sentence, which was wrong for tHN twice over.

Tests

tests/testthat/test-a60-boundary-correct-skew.R. The printed-NOTE assertion
fails on main: the old text contained "correct MLE" and "not a failure" in
the correct-skew case, and the test asserts their absence.

The tests drive the two reporting paths directly rather than through a fit.
That is deliberate, and it is the lesson from the last two PRs: the models that
reach this state sit on a flat ridge where the optimizer's stopping point is not
a stable property across platforms. #54's vcov fixture slid 9e-5 along such a
ridge and stopped triggering; #57's tolerance test passed on macOS and failed on
four other platforms. A fixture chosen for being degenerate is a fixture chosen
on a flat ridge. The reporting logic here is a deterministic function of three
recorded fields, so it is tested as one.

Coverage: both skew branches, both tHN threshold combinations, tHN under wrong
skew, exactly-one-NOTE rather than two contradicting ones, the interior and NA
cases staying silent, the convergence clause appearing at conv 52 and dropping
at 0 and NA, and the two warnings staying textually distinguishable so a
caller filtering on one does not catch the other.

Still to measure, and not claimed here: how often the new warning fires on
sound fits across models and sample sizes, a fast non-simulated-ML fixture for
an integration test, and whether NLN (the other simulated-ML model on this code
path) behaves the same. A60 measured only y_pcs_w at N = 150. That sweep is
queued on the Windows box; this PR deliberately does not assert anything about
it.

🤖 Generated with Claude Code