diff --git a/openupgrade_scripts/scripts/ptplus/16.0.5.1.2/pre-migration.py b/openupgrade_scripts/scripts/ptplus/16.0.5.1.2/pre-migration.py
new file mode 100644
index 00000000000..7c66b6ccad0
--- /dev/null
+++ b/openupgrade_scripts/scripts/ptplus/16.0.5.1.2/pre-migration.py
@@ -0,0 +1,121 @@
+import logging
+
+from openupgradelib import openupgrade
+
+_logger = logging.getLogger(__name__)
+
+# Retired namespace for xmlids of records ptplus stopped shipping but the
+# instance's history still references. No module by this name ever exists, so
+# Odoo's orphan-xmlid collector (_process_end only processes REAL modules being
+# loaded) never touches it — while the full original identity stays recorded in
+# ir_model_data (visible as External ID in dev mode, queryable forever). Chosen
+# over deleting the pointer because account.account.tag has no note/comment
+# field to carry provenance: the renamed pointer IS the provenance.
+LEGACY_MODULE = "__ptplus_legacy__"
+
+
+@openupgrade.migrate()
+def migrate(env, version):
+    # WHY (archaeology, verified against ptplus 15/16/17/18/19 data + the
+    # vendor's own migration scripts):
+    #
+    # Through 15.0, ptplus classifies PT withholding (retenção na fonte) on move
+    # lines with per-payment-guide-code TAGS: "PT-RF CD <code>" (xmlid
+    # ptplus.tax_tag_rf_cd_*, 26 tags in data/account_tax_data.xml; the codes
+    # are the official AT guia-de-pagamento codes — 102 trabalho independente,
+    # 104 prediais, 2xx IRC, ... — all still officially current).
+    #
+    # At 16.0 the vendor rebuilds the whole RF system: dedicated withholding
+    # TAXES appear (account_tax_template_rfirs_* / rfirc_*), and from 17.0 the
+    # template CSV carries the guide code as FIELDS on the tax
+    # (l10n_pt_wh_tax_type / l10n_pt_wh_code / l10n_pt_wh_income_type). The
+    # vendor's own scripts (migrations/16.0.5.1.0, .5.1.2) migrate the OTHER tag
+    # families by name filter (replace_tags()/delete_tags() on "RF DM/M10/M30/
+    # IS…", clearing the m2m before the tag goes), but no filter matches
+    # "PT-RF CD" and the family is re-declared under no other xmlid: it is
+    # dropped with NO data migration. At settle, the orphan-xmlid collector
+    # tries to DELETE every tax_tag_rf_cd_* the module no longer declares; the
+    # FK-held ones fail ("bad query: DELETE FROM account_account_tag …"), the
+    # 15->16 hop goes RED and the 18 settle hits a FATAL through the same path.
+    #
+    # FIX — selective retirement (the migration script the vendor forgot):
+    #   * CD tags REFERENCED anywhere: move their ir_model_data pointer to the
+    #     retired namespace above. Record, links and provenance all survive;
+    #     the collector's condemned DELETE is never attempted.
+    #   * CD tags referenced NOWHERE: left alone — the collector deletes them
+    #     cleanly (no error, nothing lost, no zombie rows). On an install with
+    #     no old-system history the family disappears entirely.
+    #
+    # "Referenced" is decided by ROW EXISTENCE in EVERY table holding a FK to
+    # account_account_tag, discovered dynamically from the catalog — NOT by
+    # whether the collector's DELETE would fail. The difference matters: some
+    # m2m rel tables carry ON DELETE CASCADE, so a tag linked only through one
+    # of those would be deleted "successfully" while silently stripping the
+    # classification off history. Existence-based selection is immune to that.
+    #
+    # A remap onto the new structure is deliberately NOT attempted: the
+    # replacement is a field on per-rate taxes (one code -> many taxes), there
+    # is no tag-shaped target, and nothing at 16+ reads the CD tags (verified:
+    # zero references in data/migrations through 19.0; no account.tax.report
+    # line ever pointed at them). Idempotent: on a re-run the referenced ones
+    # already sit in the legacy namespace and the unused ones are gone.
+    env.cr.execute(
+        """
+        SELECT res_id
+        FROM ir_model_data
+        WHERE model = 'account.account.tag'
+          AND module = 'ptplus'
+          AND name ~ '^tax_tag_rf_cd_'
+        """
+    )
+    cd_ids = [r[0] for r in env.cr.fetchall()]
+    if not cd_ids:
+        _logger.info("ptplus PT-RF CD tags: none pending — nothing to do")
+        return
+
+    # every (table, column) with a FK into account_account_tag, straight from
+    # the catalog so any install's extra consumers are covered too.
+    # conrelid::regclass::text comes out correctly quoted when needed.
+    env.cr.execute(
+        """
+        SELECT c.conrelid::regclass::text, a.attname
+        FROM pg_constraint c
+        JOIN pg_attribute a
+          ON a.attrelid = c.conrelid
+         AND a.attnum = ANY (c.conkey)
+        WHERE c.contype = 'f'
+          AND c.confrelid = 'account_account_tag'::regclass
+        """
+    )
+    used = set()
+    for table, column in env.cr.fetchall():
+        env.cr.execute(
+            f"SELECT DISTINCT {column} FROM {table} WHERE {column} = ANY (%s)",
+            (cd_ids,),
+        )
+        used.update(r[0] for r in env.cr.fetchall())
+
+    unused = sorted(set(cd_ids) - used)
+    _logger.info(
+        "ptplus PT-RF CD tags: %s referenced -> xmlid retired to %s (record + "
+        "links + provenance kept); %s unreferenced -> left for the orphan "
+        "collector to remove cleanly. kept=%s dropped=%s",
+        len(used),
+        LEGACY_MODULE,
+        len(unused),
+        sorted(used),
+        unused,
+    )
+    if used:
+        openupgrade.logged_query(
+            env.cr,
+            """
+            UPDATE ir_model_data
+               SET module = %s, noupdate = true
+             WHERE model = 'account.account.tag'
+               AND module = 'ptplus'
+               AND name ~ '^tax_tag_rf_cd_'
+               AND res_id IN %s
+            """,
+            (LEGACY_MODULE, tuple(used)),
+        )