[jasperblues]

Port of liberation-data/drivine4j#370 (drivine4j PR liberation-data/drivine4j#387, squash commit 8d6e12c4). Only the PR's final version is ported. Its first version, a readInPlace hook, closed Oracle's session in the object manager's save, so drivine4j replaced it. Drivine for TypeScript's port is liberation-data/drivine#430.

What happened

A text function or + over a property and an operand whose SQL is long, such as toLower(trim(coalesce($p.q, m.name))), has its operands bound by shared() in a SELECT /*+ NO_MERGE */ subquery. The property was bound there as JSON and then read as a CLOB from that JSON, so drivine4j#255's _direct() read never applied. On Oracle, a text past 32 KB read this way closed the session from the second run on, given two rows or more (ORA-03113, DPY-4011 here).

Fix

The rule behind the fix: on Oracle, reading the outer row's properties beside a NO_MERGE view can close the session, but reading them inside the view is safe.

• texts(), Oracle: binds a property's VARCHAR2 and CLOB texts beside the long value, each read by the property's own path. This covers CONTAINS, STARTS WITH, ENDS WITH, split and replace. In _long mode, only the CLOB is bound, as before.

• shared_sum(), a new hook: the + site calls it, and so does the temporal binary's read (a stored value plus or minus a duration). On Postgres it is just shared(). On Oracle, a property side next to a long value is bound as three columns, each read by its own path:

  • its JSON, null past 32,767 characters (_BOUND_TEXT);
  • its kind;
  • its CLOB.

  The body's reads of the kind and the CLOB are rewritten to those columns, the way _direct() rewrites reads.

• shared_bound(): counts each binding, so _binds() (which decides when an aggregate's argument is worked out first) still sees them.

Differences from drivine4j

• binds_long(values): this is drivine4j's binds(vararg values). _binds is already taken here.

Tests

• tests/integration/test_cypher_compiled.py: drivine4j's cross-store test, with its sync copy regenerated. It runs on every store; on Oracle every statement runs three times on one session (one_session), and the session is checked after each (asked_again).
• tests/unit/test_oracle_item_property.py: drivine4j's SQL pins from OracleItemPropertyTest:
  • text functions bind the texts by the property's path;
  • sums bind the kind and CLOB columns and read them, including the object manager's a.id + '->' + b.id;
  • both label tests stay;
  • an aggregate over these shapes still works its argument out first.
• Fails before: with only the new tests on main, the Oracle case fails in both async and sync with DPY-4011 (the database closed the connection), and the 3 SQL pins fail. Neo4j and Postgres pass.
• Gate: ORACLE_TESTS=true uv run pytest -q, the whole suite, gave 5,120 passed, 298 skipped, 0 failed. ruff check, ruff format --check, pyright and unasync.py --check pass.

Not covered

• On Oracle, a list plus a stored text past 32,767 characters gives null, as in drivine4j. This is noted on shared_sum.
• The shapes drivine4j#387 leaves for their own issues behave as they do there:
  • =, <>, IN and > beside a long value;
  • an item past 32 KB of a split list, read by index.

Fixes #232