The manager holds _historical_frame across the forecast delivery — measured at ~1.2% of a target frame, so not urgent

[Polichinel]

Split out of #269, which listed as an acceptance criterion: "The historical frame is released, or not held, across the forecast delivery phase." #269's shard-loading work landed in #271; this criterion did not, and it is filed here rather than closed over.

What it is

CRAFDPostProcessorManager / UNFAOPostProcessorManager set self._historical_frame in _read and hold it through _save — across the whole forecast delivery, which is where the memory peak is.

Why it is not urgent, measured

By its own declared dimensions — 64,742 cells x 438 months = 28,356,996 rows — the historical frame is about 108 MB at one float32 column.

For comparison, one forecast target's assembled frame at 64,742 cells x 36 months x ~1000 samples is 8.68 GB. The historical frame is therefore ≈1.2% of a single target's frame, and #271 just removed about 16.8 GB per target from the same phase.

Releasing it would be a real but small saving, and it is the kind of change that touches both partners' _read/_save sequencing — more risk than 108 MB justifies on its own.

What would make this worth doing

• A delivery that fails on memory after #269: assembling a target held three copies of it, not one — 3.06x peak to 1.13x #271, where the measurement points here rather than at the shard path.
• The historical frame gaining columns — the 1.2% figure assumes a single float32 value column, and that assumption is the thing to re-check first.
• Touching _read/_save for another reason anyway, at which point releasing the frame after the historical artifact is staged is nearly free.

Scope if picked up

Release self._historical_frame once _save_historical has staged its artifact, rather than holding it to the end of _save. Both partners, byte-identical (C-33). The coverage and provenance readers that use it (views_postprocessing/{unfao,crafd}/managers/*.py around :289, :369, :413, :424, :436) all run before that point — verify that, because it is the whole question.

Register: C-101 carries the measurement. Refs #269, #271.