Honor intent.prepare in _exec_insert to restore pre-Rails-8 caching behavior

Background
----------

Before the Rails 8 adoption commit (baf9fd4 "Rails 8 support",
2025-02-23), `_exec_insert` decided whether to reuse a cached prepared
statement using `without_prepared_statement?(binds)`, which factored
in the `prepared_statements` config:

  if without_prepared_statement?(binds)
    cursor = _connection.prepare(sql)            # fresh cursor
  else
    @Statements[sql] ||= _connection.prepare(sql) # cached cursor
    cursor = @Statements[sql]
    cursor.bind_params(...)
    ...
  end

`baf9fd48` replaced that helper with the following predicate, in line
with the broader Rails 8 changes that retired `without_prepared_statement?`:

  if binds.nil? || binds.empty?
    cursor = _connection.prepare(sql)            # fresh cursor
  else
    @Statements[sql] ||= _connection.prepare(sql) # cached cursor
    ...
  end

After that change, the `prepared_statements` config is no longer part
of the cache decision in `_exec_insert`. Calls that arrive with
non-empty binds — including raw
`connection.insert("INSERT ... VALUES ('literal')", nil, "id")`, where
`sql_for_insert` appends a single `:returning_id` placeholder — go
through the cached-cursor branch on every invocation.

How it surfaced
---------------

Three things lined up:

  1. A set of pre-2018 specs that issue the same raw-SQL INSERT with
     `RETURNING ... INTO :returning_id` more than once on the same
     adapter were removed in 2017 (commit ef321a1).
  2. While those specs were absent, baf9fd4 (Rails 8 support, Feb
     2025) widened `_exec_insert`'s cached-cursor branch as described
     above.
  3. The same spec patterns were restored in 2026 as part of porting
     the trigger-based primary-key option (rsim#2615), and ran against the
     post-baf9fd48 cache decision for the first time.

Under that combination, plus `cursor_sharing = FORCE` (the adapter
default), executing the same cached cursor a second time produces a
SQL*Net half-duplex deadlock between ruby-oci8 and the Oracle 23ai
server: the server waits in `SQL*Net more data from client` while
ruby-oci8 stays in `OCIStmtExecute -> nttfprd -> read()`. See rsim#2619
for the full diagnosis and a 5-line standalone repro.

Fix
---

Reintroduce the prepared-statements check using the Rails 8.2
QueryIntent successor to `without_prepared_statement?`: `intent.prepare`.
PostgreSQL gates on the same flag (rails-3d7458fecf49
postgresql/database_statements.rb:167).

For raw SQL paths, `intent.prepare` is left as its default `false`
(set by `connection.insert` in abstract/database_statements.rb), so
the fresh-cursor path runs and the cached-cursor combination above is
not reached.

Behavior after this change:

  - `klass.create!(attrs)` (Arel insert, prepared_statements=true)
    keeps the cached path.
  - `connection.insert("INSERT ...", nil, "id")` (raw SQL) takes the
    fresh-cursor path and never reuses a cursor with a stale OCI bind
    state.
  - Connections with `prepared_statements: false` always use fresh
    cursors, matching the pre-`baf9fd48` behavior.

Tradeoff
--------

The change moves raw-SQL `connection.insert` paths to a fresh cursor
on every call, giving up the cached-cursor reuse benefit for those
calls. That tradeoff is acceptable because:

  - Issuing the exact same raw SQL string through `connection.insert`
    twice in a row is uncommon in Rails application code (model paths
    go through Arel and parameterise their bind values).
  - A SQL*Net deadlock that wedges the rest of the suite is a larger
    problem than losing per-call cursor reuse on a rarely exercised
    path.

The cached path is preserved unchanged for ORM (Arel-based) inserts,
which are the high-volume callers.

Regression spec
---------------

`spec/.../exec_insert_caching_spec.rb` issues the same
`INSERT ... NEXTVAL ... RETURNING "ID" INTO :returning_id` SQL twice
through `connection.insert`. Without the fix the second exec hits the
half-duplex deadlock and trips the spec's 5-second `Timeout`; with the
fix it completes well under 100ms. A short comment notes the
dependency on the default `cursor_sharing = force` so a future change
in defaults does not silently turn the regression test into a no-op.

Closes rsim#2619.

Author: yahonda

lib/active_record/connection_adapters/oracle_enhanced/database_statements.rb

@@ -61,8 +61,18 @@ def _exec_insert(intent, pk = nil, sequence_name = nil, returning: nil) # :nodoc
             cursor = nil
             returning_id_col = returning_id_index = nil
             with_retry do
-              if binds.nil? || binds.empty?
+              if !intent.prepare || binds.nil? || binds.empty?
                 cursor = _connection.prepare(sql)
+
+                unless binds.nil? || binds.empty?
+                  cursor.bind_params(type_casted_binds)
+
+                  if sql.include?(":returning_id")
+                    # it currently expects that returning_id comes last part of binds
+                    returning_id_index = binds.size
+                    cursor.bind_returning_param(returning_id_index, Integer)
+                  end
+                end
               else
                 unless @statements.key?(sql)
                   @statements[sql] = _connection.prepare(sql)

spec/active_record/connection_adapters/oracle_enhanced/exec_insert_caching_spec.rb

@@ -0,0 +1,33 @@
+# frozen_string_literal: true
+
+require "timeout"
+
+describe "_exec_insert prepared-statement cache gating" do
+  before(:all) do
+    ActiveRecord::Base.establish_connection(CONNECTION_PARAMS)
+    @conn = ActiveRecord::Base.lease_connection
+    @conn.create_table :test_exec_insert_caching, force: true, id: false do |t|
+      t.integer :id, null: false
+      t.string :name
+    end
+    @conn.execute "CREATE SEQUENCE test_exec_insert_caching_seq"
+  end
+
+  after(:all) do
+    @conn.execute "BEGIN EXECUTE IMMEDIATE 'DROP SEQUENCE test_exec_insert_caching_seq'; EXCEPTION WHEN OTHERS THEN NULL; END;"
+    @conn.drop_table :test_exec_insert_caching, if_exists: true
+  end
+
+  # The deadlock this regression guards against requires both the default
+  # cursor_sharing=force *and* an amd64 Oracle Database server; under :exact
+  # or against an arm64 server image the spec would silently pass without
+  # exercising the regression. See #2619.
+  it "does not deadlock when the same raw SQL connection.insert with RETURNING is issued twice" do
+    sql = "INSERT INTO test_exec_insert_caching (id, name) VALUES (test_exec_insert_caching_seq.NEXTVAL, 'alpha')"
+    Timeout.timeout(5) do
+      @conn.insert(sql, nil, "id")
+      @conn.insert(sql, nil, "id")
+    end
+    expect(@conn.select_value("SELECT COUNT(*) FROM test_exec_insert_caching")).to eq(2)
+  end
+end