Implement partial failure handling in orchestrator

Introduce a BufdirExportErrorCode enum with values: network_timeout, storage_write_failed, file_generation_failed, activity_query_failed, row_mapping_failed, attachment_missing, unknown. The BufdirExportErrorClassifier.classify(Object exception) method returns a record ({bool isFatal, BufdirExportErrorCode code, String message}). Implement the top-level finally block inside runExport as: `try { ... } catch (e, st) { await _markFailed(auditId, classifier.classify(e)); rethrow; } finally { await _ensureTerminalStatus(auditId); }` where _ensureTerminalStatus is a no-op if already terminal.

Use a try/catch per row inside the mapping loop (not around the whole mapper call) to enable row-level non-fatal capture. Store PartialFailureReport as a JSONB column partial_failures on the audit record — do not store in a separate table to keep queries simple.

Unit tests (flutter_test): test BufdirExportErrorClassifier with every known exception type and verify correct fatal/non-fatal classification. Test the 20% row-failure threshold triggers reclassification to fatal. Test that the finally block transitions audit to 'failed' even when an unexpected RuntimeException escapes all catch blocks (simulate by throwing from the mock audit service itself). Test ExportSuccess with partial failures returns 'completed_with_warnings' status.

Test PartialFailureReport serialises and deserialises without data loss. Use a recorded fake for BufdirExportAuditService to assert exactly one terminal status transition occurs regardless of failure mode. Target ≥90% branch coverage on error handling paths.