Pure-ish CSV/plaintext import engine for CRM contact lists (Stage 3 of the restructuring plan — the user's stated priority: account list import). Parses rows, applies the restructuring spec's Locked decisions, and writes only through PhoenixKitCRM.Lists — it never duplicates that context's insert/counter/broadcast logic.

Locked decisions this engine enforces

• Import always creates a NEW contact — never a merge, never "add to existing contact by matching email". Matching duplicates across contacts is a separate, later, human-reviewed comparison screen (Stage C4c), not something this engine decides on its own.
• Per-row atomicity: contact INSERT + membership INSERT happen in ONE transaction (Lists.add_new_contact_to_list/3) — a uniqueness violation rolls both back, so a failed/duplicate row never leaves an orphan contact behind.
• Idempotent re-import: importing the identical file twice creates zero new contacts on the second pass — every row lands in skipped instead of created/added (:already_in_list if the existing membership is still active, :unsubscribed if a prior membership was removed but still holds the email slot per the DB's partial unique index — see ListMember's moduledoc). A KNOWN collision is classified straight from a batched Lists.members_by_email/3 lookup (one per chunk — see run_chunk/4) and never reaches the write transaction at all, so re-importing a large duplicate file doesn't pay an INSERT-then-rollback per colliding row; only an email the batched lookup didn't already know about goes through it, where a race against a concurrent import is still caught by the DB's own unique index.

Row pipeline

Each row goes through, in order: normalize (trim + downcase, matching the email column's citext case-insensitivity) → in-file dedup prefilter (:duplicate_in_file — a second row with the same email, e.g. a shared mailbox listed under two names, never reaches the DB) → email format check (:invalid_email) → the batched members_by_email lookup (a KNOWN collision short-circuits straight to :already_in_list/:unsubscribed, no write attempted) → the write transaction for anything left, which still classifies an idx_crm_list_members_list_email violation (a race the batched lookup couldn't have seen) by looking up the existing holder's status.

name is CSV-optional per the spec; a blank name falls back to the email itself (mirrors Contact.display_name/1's own fallback) so the contact changeset's validate_required([:name]) never rejects an otherwise-valid row. locale, if present but not in a recognized format, is silently dropped rather than rejecting the row over a cosmetic field. company has no dedicated contact field yet (linking to PhoenixKitCRM.Companies is out of scope here) — it's stashed in contact.metadata["import_company"] so the comparison screen (C4c) can still surface it later instead of the data being silently discarded.

XLSX

Deliberately not supported yet. xlsxir (the library the plan asked to evaluate) is unmaintained since 2019 — not something to add as a new dependency. xlsx_reader is an actively-maintained alternative but wasn't evaluated in depth (its own dependency footprint, streaming behavior, etc. is a separate task). CSV/TXT cover the stated priority (account import); XLSX is deferred to a follow-up rather than shipped on an abandoned lib.