What does not work

Two categories: missing work that is on the roadmap, and things stock PostgreSQL with orafce cannot do at all.

Not yet — planned

Expressible on PostgreSQL, but the wire framing or translation is not written.

Likely to bite first

• Package-scale PL/SQL. Packages (state, overloading, the PL/SQL type system), BULK COLLECT / FORALL, pipelined functions, and compound triggers. Anonymous blocks, standalone routines, and simple triggers translate — but the further a body strays from that, the more likely the translation is to be wrong rather than merely rejected.
• OUT binds and SYS_REFCURSOR. Scalar and array input binds pass through. RETURNING <cols> INTO :out DML works for python-oracledb thin (INSERT / UPDATE). PL/SQL OUT parameters (BEGIN :x := … END) and SYS_REFCURSOR are not implemented; ojdbc and ODP.NET get ORA-03001 for RETURNING … INTO rather than a silent drop.
• LOB streaming. CLOB / BLOB values are delivered inline and capped at one ~64 MiB TTC packet. No TTC LOB locators, no multi-gigabyte streaming. DBMS_LOB.GETLENGTH / SUBSTR / INSTR exist as plain SQL functions.

Type-metadata gaps

• Native BINARY_FLOAT / BINARY_DOUBLE / INTERVAL describe for ojdbc and ODP.NET. Those drivers still see NUMBER / text for those types.
• MariaDB native INTERVAL wire types (182/183). Result columns stay text (python-oracledb thin against PostgreSQL gets native intervals; against MariaDB the probe expects text until a real type path exists).
• MariaDB ROWID, autonomous transactions, WHERE CURRENT OF, PRAGMA EXCEPTION_INIT. Not implemented; earlier shims that guessed a column named id or silently dropped pragma semantics were removed.
• Native TTC encodings for BINARY_FLOAT / BINARY_DOUBLE. Complete for python-oracledb thin on both backends; ojdbc / ODP.NET get NUMBER.

Other planned items

• Non-UTF-8 wire character sets. The server side is AL32UTF8 only. Single-byte request-charset transcoding (WE8ISO8859P1, WE8MSWIN1252) is bounded but unvalidated — no non-UTF-8 thin client exists to test against.
• External auth (Kerberos / wallet / proxy auth), a warm backend-connection cache, and a listener connection rate limit.
• Full Oracle implicit type conversion / NLS behaviour, autonomous transactions, XA.

Structurally impossible on stock PostgreSQL + orafce

Out of scope — the engine underneath can’t express these:

• Flashback (AS OF SCN|TIMESTAMP, VERSIONS BETWEEN) — no time-travel storage.
• MODEL clause — no spreadsheet/array computation primitive.
• Full PL/SQL packages — package state, overloading, and the PL/SQL type system are a compiler-scale effort, not a translation.
• DBMS_SCHEDULER / DBMS_JOB — no in-core job scheduler.
• Advanced Queueing (AQ) — no equivalent messaging substrate.
• XA / distributed transactions, DRCP, Application Continuity, network compression, client-side result cache — Oracle transport / session features with no PostgreSQL counterpart.

DBMS_PIPE / DBMS_ALERT come from orafce and pass through unchanged.