variance_measure_add: skip non-finite input instead of asserting

[rocketmark]

Summary

variance_measure_add contains an assert(isfinite(d[i])) inside the accumulation loop. When corrupt optical angle data reaches this function, the assert fires, crashes the process with SIGABRT, and corrupts the on-disk config.json because the file is left in a partially-written state. The next launch then fails to parse config and crashes again — a boot loop.

Dropping one non-finite sample has negligible effect on the variance estimate. Crashing and corrupting config does not.

Demonstration

Hardware trigger: a USB disturbance (cable flex, port brown-out) causes the FPGA to emit bad timestamps. The driver produces a non-finite optical angle. That angle propagates into variance_tracker_add → variance_measure_add:

BEFORE fix:
d[0] = NaN
assert(isfinite(NaN)) → fires
Process: SIGABRT
config.json: partially written, unparseable on next launch
Result: crash loop until config.json is manually deleted

AFTER fix:
d[0] = NaN
isfinite(NaN) == false → early return, measurement skipped
stderr: "[libsurvive] variance_measure_add: non-finite d[0]=nan, dropping measurement"
Process: continues normally
Result: one sample dropped, variance estimate unaffected

The assert was presumably added as a correctness guard during development. In production, it turns a transient hardware glitch into a persistent failure requiring manual intervention.

Impact

Any caller passing data derived from optical angles is exposed. variance_tracker_add in redist/variance.h is the primary path; it is called from lighthouse calibration and pose confidence tracking. On hardware that experiences any USB instability, this assert is reachable.

Change

One file, redist/variance.h. The assert is replaced with a pre-check that returns early and logs to stderr. The accumulator (meas->n, meas->sum, meas->sumSq) is left unchanged, which is correct — including a non-finite value in any of those fields would corrupt all subsequent variance calculations derived from this accumulator.

Found via

Observed in production: tracker running on embedded hardware (Raspberry Pi, USB bus under load) would enter a crash loop after a cable disturbance. coredumpctl showed SIGABRT inside variance_measure_add. Deleting config.json recovered the system; the underlying assert was the cause.