[hmgaudecker]

Stacked on #102 (asymmetric-translog).

Problem

Persons with no later observed measurement (death, attrition) are predicted through every remaining period. In a translog, their predicted covariance explodes (about 1e19 at ages 90+ in health-cognition). The QR of the missing-measurement update then has a zero R diagonal, the floored QR derivative overflows, and the where-mask turns 0 · inf into a NaN gradient. The log-likelihood stays finite.

Change

• predict_if_observed_later (in chs/likelihood.py, used by both scans): the predict step runs only for persons with an observed measurement in a later update. All others keep their last filtered states and covariance. Their predict step runs on placeholder inputs (zero states, identity covariance) and is discarded (double-where).
• kalman_update: the QR for missing measurements is built from an identity covariance. The result is discarded as before.
• debug_loglike: filtered states are NaN for the periods after a person's last observation.

The log-likelihood does not change, because the frozen updates contribute exactly zero, before and after the change.

Tests (tests/test_frozen_after_last_observation.py)

• Explosive translog with a person last observed in period 1: the gradient matches central finite differences. Before the change, it was non-finite.
• loglikeobs and the gradients at three random parameter vectors are equal to the values recorded on the unchanged code (9863a78).
• Filtered states are NaN exactly after the last observation. A gap followed by more observations stays finite.

Full suite: 870 passed.

https://claude.ai/code/session_01Y3CuwvNUhYTSpriKYSwqmq

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.07%. Comparing base (9863a78) to head (ccbf9b5).

Additional details and impacted files

@@                   Coverage Diff                   @@
##           asymmetric-translog     #104      +/-   ##
=======================================================
+ Coverage                96.05%   96.07%   +0.02%     
=======================================================
  Files                      147      148       +1     
  Lines                    14998    15089      +91     
=======================================================
+ Hits                     14406    14497      +91     
  Misses                     592      592              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.