feat(interface): accept in-memory text records

[afourniernv]

Priority

Medium

Problem

Applications that already hold text in memory must write temporary CSV or Parquet files before calling Anonymizer. That adds avoidable I/O and complicates service integrations. Callers also need a stable way to correlate completed rows with their own request or event IDs, including when an input column name collides with a generated output column.

Proposed solution

• Add typed TextRecord(id, text) and TextRecordsInput input shapes.
• Let the Python API accept TextRecordsInput alongside the existing file-backed AnonymizerInput without changing AnonymizerInput.source or its JSON schema.
• Preserve caller IDs and input order in the public result dataframe.
• Keep the CLI file-only; do not expose nested record objects as command-line flags.
• Retain existing file input behavior and validate configured ID columns.

Alternatives considered

• Temporary files: workable, but expensive and awkward for request-oriented services.
• Accepting a raw dataframe: broader and less explicit than the two fields the pipeline needs.
• Reading private trace_dataframe state: not an appropriate public correlation contract.

Compatibility

Existing AnonymizerInput(source=...) calls, generated schemas, and CLI behavior remain unchanged. The new shape is additive.