[m13m]

Two commits, roadmap only. The first brings it up to date with phases 2 and 3 — it was committed onto the #8 branch but did not reach master, since #8 merged while it was being amended. The second reframes it around the pivot.

Nothing in lib/ changes. This is the plan for review before any of it is built.

What the pivot section says

It separates three things rather than presenting the restructure as uniformly valuable:

Genuinely new

• An explicit connection state machine, so illegal operations are unrepresentable. "A resultset is half-read, do not send another command" is a real invariant that nothing currently enforces, and it stops being theoretical once cursors exist.
• A query context — a composable query value plus a compiler to SQL. The substance of the pivot.
• ssl.ex behind the transport behaviour, which also unblocks caching_sha2_password's full authentication path.

Mostly renaming

• protocol/ → wire/, client.ex → connection/connection.ex. Worth doing while the code is small; it buys vocabulary, not correctness.

Deliberately not adopted

• Repositories — this is the persistence mechanism.
• Domain events — a strict request/response protocol with no subscribers does not need them, and per-row events in a resultset would be actively harmful.
• Service layers over functions that are already one-liners.

Reordered phases

A. Finish the text resultset   ← next, and unchanged in position
B. Restructure
C. The query layer
D. Prepared statements
E. Authentication and TLS      ← can float earlier
F. Reduce db_connection to pooling
G. Session and utility commands
H. Replication

The resultset stays first for a practical reason: every layer above it needs something to read, and a query language on a driver that cannot SELECT cannot be validated end to end.

The question to settle

Does the compiler emit SQL with interpolated literals, or SQL with placeholders plus a parameter list? Placeholders need prepared statements to execute, so the query layer would land as compile-and-inspect before it becomes compile-and-execute. It is also the only version that is not an injection hole. Stated intent is placeholders — worth your explicit agreement, since it sets the order of C and D.

Dropped

The Ecto.Adapters.SQL adapter, as a destination. The pivot builds that layer directly. Still possible — the driver would satisfy one — but no longer what this aims at.

Also recorded

The current CRUD position, plainly: three letters of four, no parameter binding and no escaping helper, so values must be interpolated by hand. And that CI does not prove the login, since the workflow has no MySQL service.