refactor: decompose process_pdf_statement into pure extraction + persistence

[boscorat]

Parent issue

#194 — function purification

Problem

process_pdf_statement (statements.py:828) is a 9-parameter monolithic function that orchestrates:

• PDF parsing (creates Statement)
• 4 temp parquet file writes (BatchLines, StatementHeads, StatementLines, ChecksAndBalances)
• Error classification
• Timing measurement
• Prints to stdout/stderr on failure

The function mixes pure data construction with side-effecting I/O, making it hard to test error paths in isolation.

Proposal

Split into (a) pure extraction: extract_statement_data(pdf, config) -> ExtractionResult, (b) pure classification: classify_result(extraction) -> PdfResult, (c) side-effecting persistence: write_temp_files(result, paths). Timing and prints should be in the orchestration layer. Also convert the batch_line dict to a typed dataclass.

Code relationships

Target function:

• process_pdf_statement() — statements.py:828

Callers (will need updating):

• StatementBatch.__process_batch_sequential() — statements.py:1377
• StatementBatch.__process_batch_turbo() — statements.py:1398 (via ProcessPoolExecutor)

Related (called from process_pdf_statement):

• Statement.__init__() — statements.py:347
• _handle_parquet_write_error() — statements.py:800
• BatchLines, StatementHeads, StatementLines, ChecksAndBalances — parquet.py

Existing tests

• tests/test_statements.py::TestGoodStatements — exercises process_pdf_statement indirectly via StatementBatch
• tests/test_statements.py::TestBadStatements — exercises error paths indirectly

No direct unit tests exist for process_pdf_statement.

Siblings

• refactor: BatchMetadata dataclass to replace 13 duplicated parameters #195 — BatchMetadata dataclass
• refactor: ExtractionContext dataclass for extraction pipeline #196 — ExtractionContext dataclass
• refactor: decompose Statement constructor into pure factory + persistence #197 — Statement constructor decomposition
• refactor: stop mutating inputs in extraction pipeline #199 — Stop mutating inputs
• refactor: add optional conn parameter to build_datamart and related functions #200 — DB connection threading
• refactor: move processing out of StatementBatch constructor #201 — StatementBatch constructor
• refactor: remove dead logs and file_path parameters from pipeline functions #202 — Dead parameter cleanup
• refactor: Parquet class accepts pre-loaded data instead of reading from disk #203 — Parquet class purification
• refactor: _handle_parquet_write_error returns error dict instead of mutating #204 — _handle_parquet_write_error purification