Prediction time columns serialize inconsistently across formats

[fegue]

birdnet prediction outputs are not serialized consistently across save methods.

Problem

Two friction points:

1. Parquet

   • start_time and end_time are written as Arrow halffloat
   • Python reads them correctly
   • R Arrow reads raw float16 bit patterns instead of seconds

2. CSV

   • start_time and end_time are serialized as "HH:MM:SS.xx" strings
   • in-memory outputs use numeric seconds

Expected

Saved outputs should be predictable across formats and languages.

• Parquet should store time columns in an interoperable numeric type
• CSV should either preserve numeric seconds or be explicitly treated as a human-readable export

Observed

In memory:

arr = prediction.to_structured_array()
arr["start_time"][:3]
# array([0., 3., 6.], dtype=float16)

Parquet schema

prediction.to_arrow_table().schema
# start_time: halffloat
# end_time: halffloat

R read-back

arrow::read_parquet("test.parquet", as_data_frame = TRUE)
# start_time
# 0
# 16896
# 17920

Expected values were 0, 3, 6.

CSV output:

00:00:00.00
00:00:03.00
00:00:06.00

Likely cause

start_time / end_time inherit float16 and are passed to Arrow as pa.float16(). That produces Parquet halffloat, which is not read correctly by R Arrow.

Proposed fix

Upcast start_time and end_time to at least float32 before Arrow table creation.

Possibly also upcast confidence when it is float16, for consistency.

Why this matters

The same prediction currently becomes:

• numeric seconds in memory
• formatted strings in CSV
• correct numeric values in Python Parquet
• corrupted-looking values in R Parquet

That makes cross-language round-tripping unreliable.