Keep the relative formulation opt-in, and resize the test gate it exposed

Two findings from measuring rather than assuming, both of which argue against
turning this on by default.

1. It costs geo-referencing. test-navstate-odom-gnss-fusion's ENU->map rotation
   error grows to 5.69 deg against that test's 5.0 deg gate, where the
   absolute-only formulation stays inside it. The mechanism is plausible: that
   yaw is observed by comparing GNSS motion in ENU against odometry-derived
   motion in map, and asserting the dead-reckoned pose absolutely, however
   weakly, is what makes it well observed. A relative-only chain with a single
   anchor does not.

2. The APE gain it buys is being paid for on a configuration that is not
   recommended anyway. On the seven BotanicGarden sequences this formulation is
   better than the absolute-only one on 5 of 7 -- but with everything else held
   fixed, fusing that platform's wheel odometry AT ALL is worse than ignoring it
   on 6 of 7. Bad odometry stays bad however correctly it is weighted.

So: default off, documented as a trade with both numbers, and the guidance is to
enable it when fusing wheel odometry without estimating geo-referencing.

Separately, MAXIMUM_SE3_FINAL_ERROR moves 0.40 -> 0.45, and that one belongs to
the covariance-accumulation commit rather than to this feature. The odometry
that test simulates is noise-only -- no slip, no systematic error -- so trusting
it more than the motion model says is free accuracy there, and the old threshold
scored an over-confident implementation better than a correct one. Worst case
across the same seven cases:

  0.365  original: absolute pose fused with the LATEST INCREMENT's covariance
  0.419  that covariance corrected to the accumulated dead-reckoning one
  0.368  ...plus a relative factor per reading (restores part of the over-trust)
  0.405  ...with the single anchor that formulation actually implies

On real data the ranking inverts: the last of those is the best arm on
BotanicGarden and the first DIVERGES on 7 of 7, estimating 4.5-6.8x the true
path length. The fixture cannot discriminate a correct odometry weight from an
over-confident one, so it must not gate that axis; it still earns its place as a
regression guard on the fusion working at all, and the threshold is sized for
that. The reasoning is recorded at the constant.

colcon test: 34/34.

Author: jlblancoc

mola_state_estimation_smoother/include/mola_state_estimation_smoother/Parameters.h

@@ -219,6 +219,16 @@ class Parameters
      * A relative factor is skipped, rather than forced, when the previous
      * odometry keyframe has already been marginalized out: a factor on a
      * variable the smoother no longer holds would resurrect it as a free state.
+     *
+     * Off by default on a measured trade, not out of caution. It is the better
+     * of the two ways to fuse on real data (BotanicGarden, better APE on 5 of 7
+     * sequences than the absolute-only formulation), but it costs
+     * geo-referencing: the ENU->map rotation error in
+     * test-navstate-odom-gnss-fusion grows past that test's 5 deg gate, where
+     * the absolute-only formulation stays inside it. Asserting the dead-reckoned
+     * pose absolutely, however weakly, is apparently what makes that yaw well
+     * observed. Enable this if you fuse wheel odometry and do not estimate
+     * geo-referencing.
      */
     bool odometry_relative_factors = false;
 

mola_state_estimation_smoother/params/state-estimation-smoother.yaml

@@ -126,7 +126,21 @@ params:
   # dead-reckoning covariance to be honest, which makes it nearly uninformative;
   # the relative one states what the reading actually measured, with the
   # covariance the motion model computes for it, and nothing accumulates.
-  # Default off until measured on more than one platform.
+  # OFF by default, and the reason is a measured trade rather than caution.
+  #
+  # On BotanicGarden -- the only dataset family in our evaluation corpus that
+  # feeds wheel odometry, a skid-steer platform in dense vegetation -- this is
+  # the better of the two ways to fuse: better APE on 5 of 7 sequences than the
+  # absolute-only formulation. But test-navstate-odom-gnss-fusion shows it costs
+  # geo-referencing: the ENU->map rotation error grows past its 5 deg gate
+  # (5.69 deg), where the absolute-only formulation stays inside it. Asserting
+  # the dead-reckoned pose absolutely, however weakly, is apparently what makes
+  # that yaw well observed.
+  #
+  # And on that same platform, fusing wheel odometry AT ALL is worse than
+  # ignoring it on 6 of 7 sequences, so the APE side of the trade is being paid
+  # for on a configuration that is not recommended anyway. Enable this if you
+  # fuse wheel odometry and do not estimate geo-referencing.
   odometry_relative_factors: ${MOLA_ODOM_RELATIVE_FACTORS|false}
 
   # High-rate same-sensor decimation for IMU [seconds]. If > 0, readings

mola_state_estimation_smoother/tests/test-navstate-odom-gnss-fusion.cpp

@@ -41,7 +41,28 @@ constexpr double MOTION_ANG_WZ      = 0.5;  // rad/s
 constexpr double ODOMETRY_NOISE_XY  = 0.01;
 constexpr double ODOMETRY_NOISE_PHI = 0.1_deg;
 
-constexpr double MAXIMUM_SE3_FINAL_ERROR        = 0.40;
+// Raised from 0.40 on 2026-08-21, deliberately and with the measurements in
+// hand, because the old value gated the wrong thing.
+//
+// The odometry simulated below is noise-only: a zero-mean per-increment
+// perturbation, no slip and no systematic error. On such a source, trusting
+// wheel odometry MORE than the motion model says is free accuracy, so this test
+// scores an over-confident implementation better than a correct one. Same seven
+// cases, worst-case final_se3_error:
+//
+//   0.365  original: absolute pose fused with the LATEST INCREMENT's covariance
+//   0.419  that covariance corrected to the accumulated dead-reckoning one
+//   0.368  ...plus a relative factor per reading (restores part of the over-trust)
+//   0.405  ...with the single anchor that formulation actually implies
+//
+// On real data the ranking inverts: on the seven BotanicGarden sequences, which
+// have genuine wheel slip, the last of those is the best arm (0.87x the
+// lightweight estimator) and the first DIVERGES on 7 of 7, estimating 4.5-6.8x
+// the true path length. So this fixture cannot discriminate a correct odometry
+// weight from an over-confident one, and must not be used as the gate on that
+// axis. It still earns its place as a regression guard on the fusion working at
+// all; the threshold is sized for that.
+constexpr double MAXIMUM_SE3_FINAL_ERROR        = 0.45;
 constexpr double MAXIMUM_ENU2MAP_ROTATION_ERROR = 5.0_deg;
 
 constexpr const char* ODOMETRY_NAME = "odom";