[E12] Pooled vs clip-level echo tokens — the deferred half of the #69 ablation

[duckyquang]

Task metadata

Field	Value
Task ID	E12
Category	Evaluation
Priority	High
Complexity	Medium
Est. days	2
Split from	#69 (E11) — the half that was runnable is resolved there
Blocked by	HF gate approval (#70)
Feeds	#72 rerun of the missing-modality numbers

Description

#69 resolved the two pooling decisions that were runnable on current data (ECG: mean, forced — the parquet is pooled; echo head: attentive over linear on val). This issue carries the half that was not runnable: study-level mean pooling vs retained clip-level echo tokens into the attentive probe.

This is the highest-leverage open experiment in the repo: on pooled inputs the cross-attention block provably cannot attend across modalities (TECHNICAL 6.4, pinned by test_echo_to_ecg_attention_is_degenerate_on_tiled_ecg_tokens), so the reported fused numbers describe a concat-equivalent model. Clip-level echo tokens restore the ecg_to_echo direction for real, and #67 flags mean-pooling as a candidate cap on the echo branch.

Everything on the code side is ready and merged: build_echo_study_embeddings(..., max_clips=N) writes the clip-level manifest, the probes/evals read it, masks are threaded, and run_metadata.json records echo_tokens_are_clip_level. What is missing is only the data: rebuilding needs the gated HF shards (MITCriticalData/mimic-iv-echo-jepa-embeddings, vjepa2.1-vitl-mimic-pt-100, 2.17 GB total), and no machine we currently run on has gate approval (#70, checked directly — content downloads 401).

Steps:

1. Someone with HF gate approval pulls the vjepa2.1-vitl-mimic-pt-100 shards (2.17 GB).
2. Rebuild the joined manifest with --max-clips (pick N from the n_echo_clips distribution already in the manifest metadata).
3. Retrain the four probes on the clip-level manifest (same seed/splits; ~minutes on CPU) and compare pooled vs clip on val.
4. Confirm ecg_to_echo attention-weight spread goes non-zero on real data, the way the [M11] Fused probe has no cross-modal interaction — one direction still dead #72 synthetic check showed.
5. If clip wins on val: rerun the missing-modality eval ([E07] Canonical missing-modality rerun with bootstrap CIs #64's script) and restate the canonical numbers; note on [E10] Diagnose fused vs single-modality baseline gap #67 that the pooling cap was real.

Acceptance criteria:

• Pooled-vs-clip val comparison with numbers, logged to logs/pooling_ablation.json next to the decisions [E11] Pooling ablation — HuBERT-ECG and EchoJEPA clip pooling #69 recorded.
• run_metadata.json provenance shows echo_tokens_are_clip_level: true for the clip run, plus manifest/checkpoint SHA-256.
• A decision on the canonical regime, and if it changes, the README/TECHNICAL numbers restated from the new run.