Converge

Converge

Every other LAD stage looks at one feature. This one looks at the whole thing and asks a single question: does the codebase still match what the project claims about itself?

Drift is not a failure of discipline, it is what happens by default. Features get abandoned halfway, decisions get quietly superseded, documentation describes the version before last. None of that shows up as a failing test. It shows up as a session — human or agent — confidently acting on something that stopped being true months ago.

This stage reports. It does not fix. Fixing without deciding what is worth fixing is how a review turns into an unplanned refactor.

1. Collect what the project claims

Whichever of these exist:

• <docs>/decisions/ records and <docs>/DECISIONS.md
• .codebase-memory/adr.md or any other ADR file
• STATUS.md — particularly anything it describes as current
• README.md and user-facing docs, for claimed behaviour and usage examples
• <docs>/<slug>/plan.md files still carrying unticked boxes
• MAINTENANCE_REGISTRY.md deferrals

If the project records nothing, say so and stop. There is no drift to measure against nothing, and the useful output is "start recording decisions", not a fabricated audit.

2. Check each claim against the code

For each substantive claim, verify rather than assume. A code-graph tool is the right instrument here if the project has one — but check it is not stale first. A graph indexed before a large refactor reports deleted symbols with total confidence, which is worse than no graph. Re-index, or fall back to Grep and Read.

Look for:

• Decisions that no longer hold. A record says the system does X; the code does Y. Either the decision was superseded without a record, or the implementation drifted from it. Both matter, and which one it is changes what you do about it.
• Documented behaviour that is not real. Usage examples that would fail if run, CLI flags that no longer exist, API signatures that changed. Run the examples where you can — an example nobody has executed since it was written is usually wrong.
• Abandoned work. Unticked plan boxes with no corresponding code; code with no corresponding plan; TODO, FIXME, NotImplementedError, pass in non-abstract methods.
• Orphans. Modules nothing imports, tests nothing runs, config for removed features, fixtures pointing at paths that no longer exist.
• Claims in STATUS.md that have quietly become history — the most common one, because a status file is only as current as the last person who remembered it.

3. Report, ranked by what it costs to leave alone

Rank by consequence, not by count:

• Actively misleading — a decision record or doc that would cause a future session to do the wrong thing confidently. This is the worst category and usually the smallest. Fix these first.
• Unfinished — real work started and abandoned. Say what state it is in and what finishing would take, so it can be scheduled or deliberately dropped.
• Stale but harmless — out-of-date detail nobody acts on. Note it; do not spend a session on it.

For each item: what the project claims, what the code does, and a one-line note on which is probably right. Be honest when you cannot tell — "the record says X, the code does Y, I cannot tell which was intended" is a genuinely useful finding and an invitation for the user to decide.

4. Hand back decisions, not edits

End with a short list of what is worth doing, and ask. Typical outcomes:

• Correct a decision record or ADR that has gone wrong — cheap and high value
• Finish, or formally drop, an abandoned piece of work
• Update STATUS.md so it describes the present again
• Feed a cluster of related problems into /lad:maintenance-session or /lad:test-quality
• Note something as a known limitation and move on — a legitimate answer

Do not start fixing during this stage. The point is to make the state of things visible; deciding what to do about it is the user's, and it is usually a different session's work.