Embedding: within-subject learning trajectory

[hanhou]

Characterize how a subject's session embedding drifts systematically over training.

Done when

• Within-subject across-session trajectory analysis for Fig 4b

[hanhou]

🔴 Not started

No within-subject across-session embedding-trajectory analysis exists (Fig 4b).

Partially enabled: embedding_space_analysis.py already extracts and normalizes session dates per (subject, session) and builds session-embedding plots — the scaffolding for ordering a subject's sessions in time is there. What's missing is the trajectory/drift analysis itself.

Related evidence: the synthetic study 04 explicitly models within-subject drift across sessions (Stage-2/2b, incl. strong non-monotonic drift with a tail held-out split) and the embeddings track it — so the real-mouse version has a validated synthetic analogue to check against.

Sensible to do together with #27 (same tool, same checkpoint, same figure panel family — Fig 4a/4b/4c).