// Run from the project root with the updated native development runtime.
// The included fixture stays small. Replace its path with your larger CSV;
// collection then switches storage automatically when its RAM budget fills.
import aner.data
import aner.units

let proposed = data.infer_csv("examples/fixtures/data_modern.csv", "NA", 8388608, 0)
print(data.schema_description(proposed))
let source = data.scan_csv("examples/fixtures/data_modern.csv", proposed, "NA")
let normalized = data.convert_unit(source, "Temperature_DegC", "K")

// OS temporary storage, at most 1 GiB of private encoded backing storage.
// A directory path may select another existing writable volume.
let managed = data.with_spill(normalized, directory: "", max_bytes: 1073741824)
let table = data.collect(managed, max_rows: 0, max_bytes: 131072, batch_rows: 64)
print(data.storage(table))
print(data.spilled(table))
print(data.spill_bytes(table))
print(data.peak_bytes(table))

// Getters preserve types, missingness, exact datetime values and current units.
print(data.text(table, 0, "ObservationID"))
print(data.datetime_ns(table, 0, "RecordedAt"))
print(data.unit_value(table, 0, "Temperature_DegC", "degC").value())

// This plan retains the immutable table, including any disk backing.
// It does not reopen the CSV and supports bounded summaries and CSV export.
let snapshot = data.scan_table(table)
let clean = data.drop_missing(snapshot)
let report = data.summarize(clean, max_bytes: 131072, batch_rows: 64)
print(data.rows(report))
print(data.unit_mean(report, "Temperature_DegC", "degC").value())
print(data.provenance(table))
print(data.provenance(report))

// Fresh path required for a durable CSV export:
// let saved = data.write_csv(clean, "spill-clean-sensors.csv", 131072, 64)
// print(data.provenance(saved))
// Explicit opt-out keeps strict RAM admission when that is required:
// let resident = data.collect(data.without_spill(source), 0, 8388608, 64)