Roadmap

What to build next

The shell already answers queries. The next code is the 24-byte triple key. After that, each feature is one check command.

scripts/status prints one line per slice, grouped by feature. A slice counts only after its tests pass with the DISABLED_ prefix removed. The shell smoke test and the explain smoke test already pass and are not part of the percentage. A fresh tree is 0 of 46.

scripts/check key builds and runs the triple-key slices, including tests that are still disabled, then prints the next id. scripts/check 0.2 is the same slice by id. The concurrency feature uses the ThreadSanitizer preset.

Already in place

Do not rebuild these. They are the baseline the features sit on.

The shell, multi-session workers, and every command. Unbuilt commands print not built yet (PLAN x.y) and do not crash.
A query may span lines, including a PREFIX before the { } group. It runs when the braces close.
MemDictionary, MemStore, and the raptor2 loader.
Lexer, parser, and binder for the supported SPARQL subset.
Left-deep planner, triple scan, nested-loop join, filter, and projection.
The page frame, including an 8-byte pageLSN at offset 0.
LogRecord factories and Transaction accessors. Serialization and the managers still throw.
Query files, the isolation specs, and the crash harness. The harness self-tests pass. The real lock manager and the real store do not, yet.

Features

Feature	Check	Done when
Triple key	`scripts/check key`	The 24-byte key round-trips and the prefix bounds match
Buffer pool	`scripts/check buffer`	Pinned frames return to zero
B+ tree	`scripts/check index`	Random insert matches a sorted vector
RDF storage	`scripts/check storage`	A clean restart sees the triples
Query execution	`scripts/check execution`	`\set join` selects a real operator
Optimizer	`scripts/check optimizer`	A rewritten plan returns the same rows
Bulk load	`scripts/check load`	About 100k LUBM triples load under a small pool
Concurrency	`scripts/check concurrency`	The feature is green under TSan
Recovery	`scripts/check recovery`	`scripts/crashtest --seeds 200` exits 0

Slice ids such as 1.4 and 8.6 stay in the test names and in not built yet (PLAN x.y). The feature name is the one you pass to scripts/check.

Left open

These have no stubs. Write the design next to the code, then a benchmark, then a writeup.

Another eviction policy beside LRU-K, compared in benchmarks/eviction/.
A different permutation set, measured in benchmarks/indexing/.
MVCC beside strict two-phase locking.
Leaf compression with delta-encoded keys.
RDFS subClassOf materialization.
A query window that sends the same text the shell sends.

How to read ahead

The notes follow the feature order. You can read WAL and ARIES before the disk exists. You cannot check slice 8.3 until the buffer pool flushes the log before the page write.

When a page says Next, that is the slice to point scripts/check at. Delete DISABLED_ only after the test is honest. A disabled test that would have passed does not move scripts/status.