[Tracking] Push pandas to the seams — execution checklist

[Polichinel]

Tracking issue for the Push pandas to the seams epic (#85). Pick off one site at a time, each its own PR with a parity proof; check off as merged.

Execution order

• S1 S1 — frames.py: generalize to_prediction_frame to S>1 (pandas→frames) #86 — frames.py → S>1 PredictionFrame [UNILATERAL] (keystone; do first)
• S2 S2 — extraction.py: frame-native sibling readers #87 — extraction.py frame-native sibling readers [UNILATERAL] (dep S1)
• S3 S3 — forecast convert-at-the-door (internal PredictionFrame) #88 — forecast "convert at the door" [UNILATERAL] (dep S1, S2)
• S4 S4 — enrichment.py: numpy/pyarrow keyed gather (drop pandas merge) #89 — enrichment.py numpy/pyarrow gather [UNILATERAL, low pri] (independent)
• S5 S5 — build_gaul_lookup.py: pyarrow-native build #90 — build_gaul_lookup.py pyarrow-native [UNILATERAL, lowest pri] (independent)
• S6 S6 — forecast outbound → arrow sample-frame wire (ships (N,S) to FAO) #91 — forecast outbound → arrow sample-frame [GATED: faoapi #45 + C-40] (dep S3)
• S7 S7 — historical inbound read_dataframe/PGMDataset retirement #92 — historical inbound retirement [GATED: pipeline-core C-40] (do not move yet)
• Close epic [EPIC] Push pandas to the seams: views-frames migration for sample-bearing predictions #85

Sequence

S1 ──► S2 ──► S3 ──► S6 (gated: faoapi #45)
        S4, S5 (independent, low/lowest priority)
S7 (gated: pipeline-core C-40) — never blocks samples

Notes

• Unilateral arc = S1→S2→S3. Landing these makes vpp's interior carry (N,S) with byte-identical S=1 deliveries — the samples-ready milestone behind frozen wires.
• S6 ships samples to FAO but needs the coordinated faoapi wire change (FAO delivery path is point/DataFrame-based — won't carry rusty_bucket's pooled draws #45). S7 is pipeline-core-gated (C-40) and never blocks the sample work.
• Parity pattern for every step: tests/test_views_frames_conformance.py (build pandas → convert → assert arrays equal) + head-to-head old-vs-new fixtures.
• Each story = its own PR → development, usual ritual (ruff/tests/CI; manual issue close on development merges).
• Cross-refs: register C-40, FAO delivery path is point/DataFrame-based — won't carry rusty_bucket's pooled draws #45.