Don't throw when writing back a keyframe the fixed-lag smoother marginalized

process_pending_gtsam_updates()'s write-back loop iterated every keyframe in
state_.last_estimated_states and read its T/V/W variables out of the
optimizer's Values unconditionally. This is a fixed-lag smoother: once a
keyframe leaves the lag window it is marginalized out and its variables go
with it, so at() threw

    Attempting to at the key "tNNN", which does not exist in the Values.

The caller treats that as fatal ("Discarding incoming observations: a fatal
error ocurred above"), so the run stopped mid-sequence and dropped every
later observation while still exiting 0 -- leaving a short trajectory whose
APE looked good because it covered only the easy opening.

Seen on two datasets, and it needs a sequence long enough to fill the lag
window, which is why no keyed CI dataset trips it:

  - Newer College 2020 01_short_experiment: key t3478, ~14% in, 11% GT
    coverage and a flattering 0.26 m APE.
  - GEODE Offroad1_gamma: key t335, ~7.5% in, 295 of 4049 scans and a
    flattering 0.383 m.

Skipping such an entry is the correct behavior rather than merely the safe
one: it already holds the last estimate the smoother produced before
marginalizing, which is that keyframe's final value. There is nothing newer
to write back.

Offroad1_gamma now runs to completion: 4039 poses instead of 295, 99.9% GT
coverage instead of 7%, est/gt path ratio 1.007, and an honest 1.303 m APE
against the beta arm's 1.282 m on the same trajectory.

Author: jlblancoc

mola_state_estimation_smoother/src/StateEstimationSmoother.cpp

@@ -2184,6 +2184,25 @@ void StateEstimationSmoother::process_pending_gtsam_updates_locked()
         const auto tleWriteback = mola::ProfilerEntry(profiler_, "process_pending.writeback");
         for (auto& [kfIdx, kf] : state_.last_estimated_states)
         {
+            // This is a FIXED-LAG smoother: once a keyframe leaves the lag
+            // window it is marginalized out, and its T/V/W variables leave
+            // optValues with it. Reading them unconditionally threw
+            // "Attempting to at the key tNNN, which does not exist in the
+            // Values", and since the caller treats that as fatal, the whole
+            // run stopped mid-sequence and discarded every later
+            // observation. Seen on Newer College 2020 (t3478, ~14% in) and
+            // on GEODE Offroad1_gamma (t335, ~7.5% in).
+            //
+            // Skipping is the right behavior rather than merely the safe
+            // one: such an entry already holds the last estimate the
+            // smoother produced for it before marginalizing, which IS that
+            // keyframe's final value. There is nothing newer to write.
+            if (!optValues.exists(T(kfIdx)) || !optValues.exists(V(kfIdx)) ||
+                !optValues.exists(W(kfIdx)))
+            {
+                continue;
+            }
+
             const auto pose = optValues.at<gtsam::Pose3>(T(kfIdx));
             const auto linV = optValues.at<gtsam::Vector3>(V(kfIdx));
             const auto angV = optValues.at<gtsam::Vector3>(W(kfIdx));