[davidhbernstein]

ttsfm() now names the standard errors it could not compute

On a fit whose Hessian is indefinite, the negative-variance coordinates lose
their standard error. That part is correct -- no standard error exists for them
at that point -- but the fit reported NaN in those cells and said nothing
at all, after minutes of work.

Measured on main (52c3bff), macOS, reading each fit's own stored Hessian.
Design is A53's: data_gen_cs(N = 300, rand = <seed>, sig_u = .5, sig_v = .2, cons = 5, beta1 = 1.5, beta2 = 2, a = 5, mu = .1), model_name = "TTHN".

seed  secs  logLik      min eig      max eig   neg eig   neg diag(H^-1)   finite SEs   warnings
3049  395   -214.2819   -0.00209627   24862       1          3 of 6          3 of 6       NONE
3002  428   -190.4871   -0.00311802   30578.7     1          2 of 6          4 of 6       NONE

Seed 3002 reproduces the row A53 already recorded (-190.4871, -0.003118,
4/6) to every printed digit, so that measurement now holds on a second platform
and a later commit. The NaN count equals the negative-diagonal count
parameter for parameter.

The change

.se_missing_report() (R/matrix_utils.R) names the parameters whose standard
error is missing, in the fit's own reported names, and reports the Hessian's
smallest and largest eigenvalue. ttsfm()'s "TTNE" and "TTHN" branches
warn with it and record it as $se_unresolved, which print() and
summary() also report -- a warning does not survive saveRDS(), which is the
same reason .sfa_report_boundary() exists.

The message does not pick a reading. An indefinite Hessian here is either
a point that is not a maximum or a boundary solution, where the Hessian in the
full parameter space need not be positive definite. A53's item 6 is explicit
that seeds 3002 and 3049 are not the same phenomenon -- 3002's sigma_v has
collapsed onto its bound, 3049's is interior and merely unpinned -- so the
warning states both and points at tt_boundary_report() for the boundary case
and vcov(type = "bhhh") for a covariance that needs no Hessian. Three
wordings: indefinite, positive-definite-but-near-singular, and "the Hessian
could not be read", the last so that no claim is made about eigenvalues that
were never computed.

solve() on a singular Hessian is now caught. It was unguarded, so an
exactly singular Hessian threw away a fit that had already taken minutes;
psfm() and selsfm_kts() already guarded theirs. It now reports the standard
errors as missing, which the new warning then explains.

What this does NOT do

A53's items 3 and 4 -- making the substituted and unsubstituted paths report
the same Hessian estimator -- are untouched and still need David's call,
because they change the standard errors existing fits report. This PR only
makes the fit say what it could not compute. Item 5 is refuted and recorded as
such (R's sqrt() is element-wise, so there was never anything to implement).
The helper is model-agnostic; the other entry points that return NaN
standard errors silently could call it next, and that is deliberately not in
this PR.

Tests

tests/testthat/test-a53-missing-se-report.R, 30 assertions. The mechanism is
driven on hand-built Hessians rather than through a fit, because the fits that
reach this state sit on flat ridges -- which is exactly what does not reproduce
across platforms (the lesson from #54, #57 and #59). Covered: nothing reported
when every standard error is finite; the right parameters named with the right
eigenvalues; both readings present in the message; the unreadable-Hessian and
near-singular wordings; positional fallback labels; and the print() NOTE from
a constructed fit. The live reproduction is a seed-3002 fit behind
SFA_TEST_SLOW, since one TTHN fit is ~400s.

Local suite

Branch build, macOS, load_package = "installed", NOT_CRAN=true,
TESTTHAT_CPUS=2: pass 5517, fail 0, warn 48, skip 9 (the skip is the
SFA_TEST_SLOW fit).

Worth noting against expectation: the warning count is unchanged from the
A64 branch's run of the same suite (also 48). Adding a warning did not add one
here, because no fit in the suite reaches an indefinite TTHN Hessian -- the
state needs ~400s fits on specific seeds. So this change adds no noise to the
existing suite. No same-invocation main baseline was taken this round
(the machine was at load 32 throughout), so 48 is reported, not compared --
though two independent branches agreeing at 48 is itself suggestive.

🤖 Generated with Claude Code