macrame-db 0.7.0

A Bitemporal Graph Ledger on libSQL · Embedded knowledge database

Macrame

A Bitemporal Graph Ledger on libSQL · Embedded knowledge database

Published as macrame-db because macrame was already taken on crates.io by an unrelated crate. The import path is unchanged:

[dependencies]
macrame-db = "0.7"

use macrame::prelude::*;

The Python package resolves the same way, deliberately — distribution macrame-db, import macrame:

pip install macrame-db

import macrame

One caveat the Rust side does not have. A crate's [lib] name is namespaced per build graph, so macrame-db providing macrame collides with nothing. site-packages is flat. The PyPI package named macrame is an unrelated, effectively abandoned build tool (0.0.1, 2021) — but if it also installs a top-level macrame/, then installing both leaves two distributions contending for one directory, and which wins depends on install order. pip warns on file conflicts, so this is a known and non-silent risk rather than something designed around. Importing as macrame_db is the fallback if it ever matters.

Macrame is a domain-specific embedded database layer for a knowledge-ledger application: a system in which concepts are linked by typed, weighted relationships, both concepts and relationships change over time, and the history of those changes is itself a first-class asset.

Delivered as a single Rust crate that an application links directly. The entire database is one file on the local filesystem — no server, no network protocol, no external service.

Core Stack

• libSQL (MIT, unmodified) — engine with WAL, F32_BLOB, DiskANN, JSON1, window functions
• Rust, async — tokio runtime, safe Rust above the engine boundary
• Target platform — Windows desktop, embedded, single-file

Five Capabilities

Capability	Mechanism
Graph storage & traversal	Recursive CTEs over relational edge tables, compiled from a typed builder
Bitemporal semantics	Two independent clocks per row — valid time and transaction time — enforced by engine triggers
Native vector search	Per-model F32_BLOB tables with auto-maintained DiskANN indexes; vector_top_k + vector_distance_cos
In-memory graph analytics	Dijkstra, A*, SCC, k-core, Louvain (phase-one) — native adjacency-list Subgraph, no external graph dependency. Runs on a Subgraph whose adjacency is closed over its node set, so a retired concept removes its edges rather than leaving the algorithms to disagree about a dangling one
Point-in-time reconstruction	Append-only transaction_log folded with window functions; snapshot composition for fast replay (carve-outs: off across archive boundary, no cadence yet)

Two Semantic Operations

The distinction between these two runs through the entire design:

• as_of(ts) — valid-time question answered under current belief. Reports what the world looked like at ts given everything we know now, including corrections recorded after ts. A filtered read of live tables — cheap.
• reconstruct(ts) — transaction-time question. Replays the log and reports what the database actually believed at ts, before later corrections arrived. A fold over history — costs what history costs.

Both are correct answers to different questions. Conflating them is a defect.

Eight Doctrine Invariants

Every design decision derives from these invariants:

1. The boundary is sacred — Everything above libSQL is ours; everything below it is upstream. Never patch the engine.
2. Two clocks, never mixed — Valid time and transaction time are independent. No trigger or default may derive one from the other.
3. Assertions are immutable — Rows in links are never updated in place. The past is never rewritten; it is only ever superseded.
4. The ledger is a table, not the log — Transaction-time reconstruction reads transaction_log, not WAL or CDC frames.
5. No physical deletion in hot tables — Rows leave through the archive path only. Ad-hoc DELETE aborts at the trigger layer.
6. Derivative state is disposable — links_current is a rebuildable materialization. Drift is detectable by audit, recoverable by rebuild.
7. Embeddings are immutable per version, excluded from the ledger — Vectors live in per-model tables; they never appear in transaction_log payloads.
8. Fidelity is a parameter, never a silent default — as_of(ts) and reconstruct(ts) say what they mean in their signatures.

Architecture at a Glance

Application (Rust, async)
│  typed API — no SQL visible
▼
┌──────────────────────────────────────────┐
│              macrame crate               │
│                                          │
│  schema/  graph/  temporal/  vector/     │
│  DDL      CTE     as_of     DiskANN      │
│  triggers replay  snapshot  top-k        │
│  migratns algorithms archive  per-model   │
│                                          │
│         connection.rs                    │
│  (Write Actor, priority channels, clock) │
└──────────────────┬───────────────────────┘
│  libsql crate — dependency, never a fork
▼
libSQL engine (MIT, unmodified)
├── transaction_log  (append-only, trigger-captured)
├── macrame.db       (hot: current belief, recent log)
└── macrame_archive.db (cold: superseded history)

Concurrency Model

One process, one writer, many readers under WAL journaling:

• Write Actor — the sole write-capable connection lives inside a dedicated Tokio task. No other code path can name it.
• Two-tier priority channels — high-priority (UI-driven work) preempts low-priority (background jobs) at every transaction boundary via biased select!.
• Cooperative chunking — low-priority workers chunk at 500–1,000 rows, bounding lock hold to 2–3 ms per chunk.
• Read guard — the read connection carries PRAGMA query_only = ON, converting Rust ownership into runtime enforcement.

Schema

Table	Role
concepts	Entities with mutable attributes; updated in place, history in log
links	Full bitemporal assertion history; 5-column PK including recorded_at
links_current	Trigger-maintained materialization of current belief; traversals read only this table
transaction_log	Append-only ledger captured by engine triggers; the sole transaction-time mechanism
analytics_annotations	Derivative table for algorithm results; disposable, excluded from ledger
embeddings_<model>	Per-model vector tables with DiskANN indexes; created by register_model()

All temporal columns use a canonical 27-character timestamp form: YYYY-MM-DDTHH:MM:SS.ffffffZ, enforced by CHECK constraints.

Crate Layout

src/
├── lib.rs                  # public re-exports, prelude
├── error.rs                # DbError (thiserror)
├── connection.rs           # Database handle, Write Actor, priority channels, clock
├── schema/
│   ├── ddl.rs              # all DDL as const strings
│   ├── migrations.rs       # user_version-driven runner
│   └── seed.rs             # optional bootstrap
├── graph/
│   ├── builder.rs          # Traversal builder → CTE; AttributeMode hydration
│   ├── edge.rs             # assert / retire / re-assert lifecycle
│   ├── vector_filter.rs    # strategies, byte-budget cost model
│   ├── subgraph.rs         # DB → Subgraph loader, byte budget
│   └── algorithms.rs       # dijkstra · astar · scc · k_core · louvain
├── temporal/
│   ├── interval.rs         # Interval, overlap arithmetic
│   ├── as_of.rs            # valid-time filters under current belief
│   ├── replay.rs           # window-function reconstruction; cold-DB ATTACH
│   ├── snapshot.rs         # bincode + zstd snapshots, seq-anchored
│   └── archive.rs          # ATTACH-based cold storage
├── vector/
│   ├── embedding.rs        # Vec ↔ F32_BLOB codec
│   ├── model.rs            # ModelName newtype (validated identifier)
│   ├── registry.rs         # register_model, declared_dimension
│   └── search.rs           # top-k, RRF fusion
├── integrity/
│   ├── audit.rs            # audit_current() — read-side
│   └── rebuild.rs          # rebuild_current() — high-priority command
└── util/
    ├── ids.rs              # ULID generation & validation
    └── clock.rs            # Clock trait; SystemClock; FakeClock

Python Bindings

The crate is the product; the binding is a second front door onto the same ledger. Everything below runs against the same engine, the same schema and the same Write Actor — no reimplementation, no second source of truth.

pip install macrame-db

import macrame

T0 = "2026-01-01T00:00:00.000000Z"

with macrame.Database.open("kb.db") as db:
    db.write_concepts([
        macrame.ConceptUpsert("bitemporal", "Bitemporal Modelling", valid_from=T0),
        macrame.ConceptUpsert("valid-time", "Valid Time", valid_from=T0),
    ])
    # Both endpoints must exist: `links` carries a foreign key onto `concepts`.
    db.assert_edge(
        macrame.EdgeAssertion("bitemporal", "valid-time", "CITES", valid_from=T0)
    )

    graph = db.load_subgraph("bitemporal", 3, 1 << 20)
    print(graph.dijkstra("bitemporal"))

Use the context manager. close() is the only path that writes the final snapshot and the only one that can report the write actor's exit status. A handle that is merely garbage-collected loses both, at a time Python chooses.

What crosses the boundary

Writes	assert_edge, retire_edge, upsert_concept, write_concepts, bulk_import, write_bulk_atomic, annotations
Reads	traverse, traverse_ids, load_subgraph → an opaque Subgraph handle
Analytics	dijkstra, astar, scc, k_core, louvain, modularity — on the Subgraph, not on a copy of it
Temporal	reconstruct, archive, archive_windowed, verify_snapshot_chain, query_as_of_edges
Vector	register_model, upsert_embeddings, search_vector, keyword_search, hybrid_search, search_filtered
Integrity	audit_current, rebuild_current, rebuild_current_chunked, rebuild_fts
Introspection	metrics() — real counters, because the wheel is built with --features metrics unconditionally. A Rust consumer can turn a feature on; a Python one cannot

The API is synchronous. The tokio runtime is process-global and lives behind the boundary; every call releases the GIL for its duration, so other Python threads run while a write is in flight. astar is the one exception — it holds the GIL, because its heuristic is a Python callable and re-attaching once per node expansion would cost more than releasing saves.

Where Python differs, and why

• Timestamps are aware datetime objects or canonical YYYY-MM-DDTHH:MM:SS.ffffffZ strings. Naive datetimes are refused rather than assumed to be UTC.

• An open interval is None, in both directions — never a sentinel datetime. The stored sentinel is exactly datetime.max, and .astimezone() overflows on it for any zone east of UTC.

• Values validate in their constructor. A malformed EdgeAssertion fails on the line that built it, not on the write that consumed ten thousand of them with no index to say which.

• Every error is typed and carries its fields as attributes — 35 exception classes under MacrameError, with six intermediate groups (IntegrityError, ValidationError, VectorError, TemporalError, WriterError, BudgetError) that exist to be caught as sets:

  try:
      db.assert_edge(edge)
  except macrame.OverlappingIntervalError as e:
      print(e.source_id, e.valid_from, e.existing_from)
  

• Subgraph is an opaque handle, not a dict. It is the one thing the ledger materialises under a byte budget, so it answers questions about itself rather than copying itself into Python. .to_dict() is the explicit purchase of the copy.

• Deliberately not exposed: the TraversalBuilder (Python gets keyword arguments), AttributeMode.OMIT on traversals, and raw() — an escape hatch whose safety is a Rust convention that does not survive the crossing. diagnostic_query() is the read-only replacement.

Distribution

One abi3 wheel per platform, CPython 3.10+: the extension links only the stable ABI, so a single wheel serves 3.10 through whatever ships next instead of one build per minor version. An sdist is the fallback for platforms the matrix does not cover — it builds from source and needs a Rust toolchain. Type stubs ship inside the wheel and py.typed is set, so a checker sees the whole signature, including which arguments take an aware datetime and which return one.

One process, one open handle per database path. Not a style preference: R15 (below) faults on concurrent open of local libSQL databases, and it reproduces through this binding at the same rate as through the crate. Use threads — the GIL release is what makes them worth having. If you need processes, use the spawn start method; a fork() guard poisons inherited handles rather than letting a child write to a database no actor is serialising.

Not on PyPI yet. The wheel matrix and the publish workflow exist and have never run, so the two PyPI badges above will read not found until the first release is uploaded. Until then, pip install . from a checkout goes through the same maturin backend the wheels do.

Full reasoning: §14 of the architecture set and the bindings plan.

Testing

# Full test suite (unit, integration, scenario)
cargo test --no-fail-fast

# Property tests (generated-history binaries, run serially)
cargo test --features property-tests --no-fail-fast

# The Python suite — through the gate, never bare pytest
python tests_py/run_suite.py

--no-fail-fast is not optional for reading the totals: without it cargo stops at the first failing binary and everything alphabetically behind it never runs. That is how the archive defect U sat unnoticed behind a known-red concurrency_tests.

run_suite.py is not a convenience wrapper. R15 kills the interpreter rather than raising, and the two ways that reports — no summary at all mid-run, or a green summary with a non-zero exit when the fault lands in Drop — are not both caught by any single signal. It checks the summary, the counts and the exit code against each other, names four outcomes, and retries only the crash.

Test Layers

Layer	What it proves
Unit tests	CTE builder output, interval arithmetic, RRF fusion, embedding codec roundtrips. Note that the interval-arithmetic test is the only caller of Interval::overlaps — defect AG
Integration tests	Full API against real database files; WAL crash recovery
Property tests	Random assertion/retirement streams never produce overlapping open intervals; links_current == latest-belief projection. The open-interval restriction is load-bearing, not incidental — nothing tests the closed case, which is defect AA
Scenario tests	Attribute fidelity across AttributeMode values; corrupt-then-rebuild roundtrip
Regression tests	tests/wave1_regression_tests.rs — one per Wave 1 defect, each verified to fail against the pre-fix tree. Its header names the three that pass either way and says why, rather than letting them look like coverage they are not

Current baseline: 296 passing, 0 failing across 26 binaries on plain cargo test, and 305 with --features metrics, which is how CI runs it. The Python suite is a separate 344.

Read a property-test total with care. --features property-tests most recently reported 308 passed, 0 failed with eight tests silently absent: R15 had killed a binary, and since cargo test runs one process per binary the others still printed their summaries and the aggregate looked green.

Benchmarks are separate and are measurements, not gates — §9's numbers are stated for named hardware, so an absolute threshold in CI would gate on whichever runner picked up the job:

cargo bench --bench budgets

Two things to know before reading a red or a green here. A green means "no test disagrees with the code" — the 2026-07-30 review found ten defects living inside a green suite, and while nine are now closed by tests, one still is not. And a red may not be yours: R15 crashes an arbitrary test binary with STATUS_ACCESS_VIOLATION on roughly one full run in three, taking a different binary each time and passing when that binary is re-run alone. Re-run the named binary on its own before believing it.

Known test gaps

Gap	Detail
Raw-SQL overlap	The valid-time overlap guard lives in the write actor, so it binds callers of this API and not raw SQL against the same file. A trigger would bind both and would tax every insert on the path the v6 index was just added to make fast.
Unbenched read paths	benches/budgets.rs measures twenty rows and none of them is load_subgraph, one of the five algorithms, archive(), or FilteredVectorSearch. The deferred Subgraph rewrite is blocked on a benchmark in that gap.
Benchmark gates	§9 performance budgets are measured, not enforced as CI gates — deliberately (D-047, D-055).
RecordedAtRegression	Mapped by the error classifier but unreachable through the public API — SystemClock is strictly increasing by contract.
Plan-shape coverage	Three tests now pin a query plan rather than a result (D-042, D-059, D-064), because that class of defect returns the right answer and no correctness test can see it. Every index-sensitive query should have one; not all of them do.

Known defects

Severity rule: a wrong answer outranks a crash. Full evidence, reproductions and remediation order are in the implementation plan §8.5; the letters are that document's register.

Open:

#	Location	Defect	Wave
Louvain phase two	graph/algorithms.rs	Local-moving only; no community aggregation, so coarse structure is missed. Deferred on the byte budget's ceiling, which Wave 3's measurements weakened — 28 ms at 10K nodes leaves room. Open pending a caller who loads graphs where it pays	—
Snapshot cross-check	temporal/snapshot.rs	Chains compound: write_final composes onto the previous snapshot, so an error propagates forward with nothing verifying a composed result against a fold from genesis	—
No verify_fts()	connection.rs	rebuild_fts() is the repair with no way to ask whether it is needed. FTS5's integrity-check cannot answer — it verifies the index's internal consistency, not its agreement with concepts — so none was written rather than one that would call an empty index healthy (D-071)	—
R15	libSQL 0.9.30	Intermittent STATUS_ACCESS_VIOLATION when local databases are opened concurrently in one process; mitigated by RUST_TEST_THREADS = "1" and the property-tests feature gate	—

Closed by Wave 4 (2026-07-30) — hardening, recorded as D-066 … D-070:

#	Was	Resolution
load_subgraph growth	Superlinear, 12.5× for 10× the nodes, unexplained	Explained, not fixed (D-070): an O(E log E) DISTINCT sort, load-bearing. Two fixes tried — deduplicating the walk first is 36% slower; removing the redundant second sort measures 85.95 vs 86.76 ms, i.e. nothing
ATTACH race	The cadence shared read_conn, so two folds could interleave in the unsynchronised ATTACH cold … DETACH cold region	The cadence gets its own connection (D-066) — removes the shared state rather than ordering access to it
close() swallowed failure	The writer's Result was discarded, so a Database whose actor had panicked closed "successfully"	Returns DbError::WriterStopped, before the final snapshot is written
Stale snapshots after upgrade	A SCHEMA_VERSION bump invalidates every snapshot and nothing wrote a replacement, so the first reconstruct folded from genesis	Re-anchored at open (D-067), gated on it being a real upgrade and the cadence being enabled
Errors naming the wrong subject	timestamp::normalize reported caller input as ReplayCorrupt { seq: 0 }; archive_horizon.archived_at was written with the cutoff	DbError::InvalidTimestamp; the archive time comes from the clock (D-069)

Closed by Wave 3 (2026-07-30) — the measurement wave, recorded as D-063, D-064 and D-065:

#	Was	Resolution
AI	The overlap guard's narrowing predicate made idx_lc_traversal_cover win as a covering index, so the guard bound one column and scanned the source's out-degree — D-059's defect, reproduced by D-060's fix one wave later. +9.8 ms per 90-edge chunk into a 2,000-edge hub	Predicate dropped; the query is now a three-column point lookup. No correctness test could have caught this — the guard answered correctly throughout — so the plan is pinned by a test
AF	corpus_size runs COUNT(*) per filtered query; the planner's input is O(corpus)	Closed without a fix. Measured at <1% of a search, and caching it would be unsound: it changes on every embedding write
D-047	The Subgraph integer-index rewrite, deferred on a benchmark nobody had written	Retired. The load dominates any single analysis at every size measured

Closed by Wave 2 (2026-07-30), with three decisions recorded as D-060, D-061 and D-062:

#	Was	Resolution
AA	trg_links_single_open guards only the open sentinel, so overlapping closed intervals were accepted and read back as two edges for one relationship	Refused in the write actor (D-060). Not in a trigger — so raw SQL can still write one, and §4.2 says so
AG	Interval::overlaps was dead code, the missing half of AA	It is the guard's decision procedure; the SQL beside it only narrows
AD	Nothing validated an identifier while three modules assumed ULIDs	Ids are opaque with | and / reserved (D-061). Every existing id stays valid, and W's collision becomes unreachable rather than merely harmless
J	validate_id returned NotFound for a malformed id	DbError::InvalidId { id, reason } — a refusal is not an absence
K	FakeClock was constructed in the harness and injected nowhere	Database::open_with_clock (D-062), with the floor lifted into the Clock trait
D-059	The single-open probe scanned its source's out-degree on every insert	idx_lc_open_interval on a v5 → v6 rung. SCHEMA_VERSION is now 6

Closed by Wave 1 (2026-07-30), each by a named test rather than by a commit — the distinction is deliberate, because defect H spent a full cycle marked fixed on the strength of a commit that changed nothing observable:

#	Was	Resolution
V	Concept log payload omitted embedding_model, so every temporal read returned None	Payload v2, readers accept v1 and v2
W	Folds partitioned on entity_id alone; a concept colliding with a link key vanished from the reconstruction	Partition on (table_name, entity_id)
Z	Adjacency could reference absent nodes: louvain panicked, scc emitted phantoms, k_core inflated degree	Subgraph carries a stated closure invariant; dangling entries dropped. No algorithm needed changing
AB	Three attribute readers, three retirement semantics	One rule: not returned if retired as of the instant asked about
AE	hydrate and hydrate_attributes issued one query per node	Batched at 400 ids per statement. Not yet benchmarked — Wave 3
AC	classify_archive_violation called from nowhere, so DbError::ArchiveViolation was unconstructible (reopened H)	Deleted — error::classify already did the same job. archive()'s deletes routed through it
AH	Doctrine VII's static guard banned the substring embedding, refusing the V fix	Narrowed to permit embedding_model by name, with a test pinning that it is still a scalar column

Two defects this table used to carry, S (CostEstimator with no strategy implementations) and T (RRF with no keyword arm), are fixed — D-050 and D-051 respectively. They were listed as open here while the delivery table below recorded them as delivered.

Minimum Supported Rust Version

1.88, verified rather than declared — cargo +1.88.0 check --all-features --all-targets passes and 1.85 does not.

It is worth saying where that floor comes from, because it is not this crate. Macrame's own code needs 1.73 (div_ceil; 1.70 for Option::is_some_and). The binding constraint is home@0.5.12 at 1.88, reached through libsql-ffi's build-dependency on bindgen → which → home — and a build-dependency binds consumers too, since libsql-ffi compiles the SQLite amalgamation on every downstream build. Lowering the floor means pinning that chain (cargo update home --precise …), not changing anything in this crate.

The MSRV is treated as a compatible change and may rise in a minor release.

Dependencies

Crate	Role
libsql 0.9.30	Engine binding
tokio 1	Async runtime; actor task, channels
serde / serde_json	Payload serialization
bincode	Snapshot serialization
zstd	Snapshot compression
thiserror	Error derive
tracing	Structured diagnostics
ulid	Entity ID generation

No GPL-licensed components. No chrono or time dependency.

Delivery Status

Phase	Status
Phase 0 — Schema + migrations	Delivered — legacy-free baseline at v2, rungs v2→v3 (D-041) and v3→v4 (D-042)
Phase 1 — Write Actor + public write API	Delivered — exhaustive match, no wildcard; assert/retire/upsert/bulk-atomic/bulk-import/annotations/rebuild/archive
Phase 2 — Temporal core (replay, snapshots)	Delivered — ATTACH bracketed on all paths, self-healing (D-044), versioned container (D-043)
Phase 3 — Vector + graph	Delivered — register_model + upsert_embeddings through the actor (D-048); native Subgraph with five algorithms (D-039); edge types bound, not interpolated
Phase 4 — Document restoration	Delivered — §5.2–§5.9, §6, Appendix A de-corrupted and forward-ported
Snapshot composition (D-049)	Delivered — anchored fold + tombstone merge; composes across the archive boundary as of D-052
Snapshot cadence (D-053)	Delivered — read-side maintenance task, triggered by log distance rather than a clock; close() stops and joins it before the final anchor
Snapshot retention (D-054)	Delivered — the newest five plus one per day for thirty days, as §5.5 always specified; container header v2 carries each snapshot's instant so bucketing costs 18 bytes, not a decompression
§9 benchmarks (D-055)	Delivered — cargo bench --bench budgets measures twelve budget rows. Measurement, not CI gates: absolute durations on arbitrary runners are the wrong shape (D-047's reasoning). Eleven pass; chunk commit missed by 20×
Chunk commit (D-056)	Delivered — statement prepared once per chunk instead of once per row: ≈62 → ≈37 ms at 500 rows. Measuring the residual showed the ledger triggers are ~92% of it, and that §9's ≤ 3 ms is the un-amplified cost — so §5.1.5's golden rule needs re-deriving, not the code
Bulk write paths (D-057)	Delivered — the same hoist in the three bulk paths §9 does not budget: concepts 34.1 → 11.9 ms, annotations 4.60 → 2.13 ms, embeddings 73.4 → 67.4 ms at 500 rows, measured by a new bulk_chunks group rather than assumed from the edge chunk. The spread inverts D-056's reasoning: analytics_annotations has no triggers and saved 54%, the DiskANN tables saved 8%, so preparation is a near-constant per row and the saving is largest where the row is cheapest
Superlinearity diagnosed (D-059)	Delivered — and it corrects D-058. The sweep measured chunk size and table size as one variable; separated, a fixed 90-row chunk into a hub of 0/2,000/8,000 edges costs 4.4 / 18.4 / 47.7 ms, so the cost is in the table. End to end, 1,000 edges are 85.5 ms as one transaction vs 94.7 ms as eleven chunks — chunking is ~11% slower, not 3.3× faster. The edge path's growth is a defect: the single-open guard's EXISTS plans as a one-column scan over idx_lc_traversal_cover, so every insert scans its source's out-degree (90 rows into a 90,000-edge hub: 1.06 s), and that hits interactive assert_edge too. Fix proven (47.7 → 8.0 ms, flat), not shipped — it needs a migration rung. Embeddings are the same shape and inherent to DiskANN. See examples/chunk_diag.rs
Golden rule (D-058)	Delivered — §5.1.5 re-derived from a chunk-size sweep. It is a bound on duration, so CHUNK_ROWS = 1000 becomes CHUNK_BUDGET = 3 ms with four measured sizes — edges 90, concepts 70, annotations 600, embeddings 30 — at 2.39 / 2.35 / 2.36 / 2.06 ms where the single constant gave 89 / 24 / 3.5 / 143 ms. Two paths turned out superlinear in chunk size, so smaller chunks are 3.3× faster in total there, not a trade; per-transaction overhead is ~0.8 ms, not "noise". Open: why the superlinearity
Archive read path (D-052)	Delivered — hot_log_covers tested how far the hot log reached, not whether it was complete, so a reconstruction before the archive cutoff could silently drop an entity
Phase 5 — Test matrix	Delivered — Doctrine VIII divergence (both directions), archive crash safety (D-012), Doctrine VII property suite, and empirical cost estimates via D-050
Filtered vector search	Delivered (D-050) — FilteredVectorSearch; two strategies, both with bodies, held together by a test requiring them to agree; TwoPhaseTempTable removed, its two mechanisms being absent from libSQL 0.9.30
Hybrid search	Delivered (D-051) — concepts_fts FTS5 external-content index on a v4 → v5 rung; HybridSearch fuses vector and keyword arms by RRF; rebuild_fts() satisfies D-036
Subgraph integer-index rewrite	Deferred — pending a Louvain/Dijkstra benchmark on a budget-sized graph, which does not exist; see Wave 3

Every item the previous plan sequenced is delivered. What comes next is set by the 2026-07-30 review rather than by that sequence.

Roadmap

Four waves, ordered by the project's own severity rule — a wrong answer outranks a crash, and both outrank a slow one. Full detail, evidence and per-item acceptance tests are in the implementation plan §9.

Wave	Theme	Scope	Size
1 ✅	The six silent defects	V, W, Z, AB, AC, AE, plus documentation drift. No schema change. Delivered 2026-07-30; thirteen new tests, 171 → 184 passing	done
2 ✅	Decisions deferred a full cycle	D-059's index shipped as the v5 → v6 rung; valid-time overlap (AA/AG), identity (AD/J) and the injectable clock (K) all decided. Delivered 2026-07-30; 184 → 202 passing	done
3 ✅	Measure what the bounds claim	Six bench groups covering every previously unmeasured path. Found one defect (AI), retired two proposed optimisations, confirmed the four chunk constants. Delivered 2026-07-30; 202 → 203 passing	done
4 ✅	Hardening	The ATTACH race, close()'s discarded error, raw()'s surface, the migration re-anchor, three mis-named errors, and Wave 3's unexplained superlinearity. Delivered 2026-07-30; 203 → 205 passing	done

| 5 ⬅ | The last silent-wrong-answer paths | The unreachable 'D' branch (D-072); the FTS5 VACUUM hazard (investigated, nothing built — it isn't reachable, D-071); load_subgraph_with, which found two accounting defects older than itself (D-073). Remaining: R15, one consolidated §4 statement, the rename | in progress |

Waves 1–4 are delivered; Wave 5 is under way. What remains open is in the table above, and none of it returns a wrong answer.

Wave 5's first two items are worth reading as a pair, because both ended by removing something rather than adding it: an unreachable branch and the set that served it (D-072), and two mechanisms declined after measurement showed one hazard unreachable and one check unable to see it (D-071). The output was a smaller crate and three facts nobody had written down.

Why correctness preceded the performance work: D-059's index is the largest measured win in the tree and it moves user_version, so it wants a stable base and its own migration test. Wave 1's defects needed no schema movement, so they landed while that rung was still being decided — Wave 2 is now what's next.

Contracts settled by Waves 1 and 2

Recorded here because they constrain future work rather than merely describing past work.

• Subgraph is closed. Every adjacency endpoint is a hydrated node, stated on the type. Algorithms may rely on it and none re-checks. A future loader that admits retired nodes has to revisit all five.
• Retirement is uniform. A concept retired as of the instant asked about is not returned, by any of the three readers. Current asks "retired now" and AtTime asks "retired then" — that difference is the two clocks and is meant to stay.
• One error classifier, not two. error::classify is the only path to a typed guard error; the archive's private duplicate is gone.
• Identifiers are opaque, except that | and / are reserved. Not ULIDs — the crate never required them and three modules only assumed them. The two characters delimit the transaction log's entity key and the traversal path, so an id carrying one is ambiguous rather than merely unusual. generate_id() is offered, not required.
• Valid-time intervals for one relationship never overlap — through this API. Two open intervals are SingleOpenViolation (the trigger), anything else overlapping is OverlappingInterval (the write actor). The two guards partition the space; they do not layer. Raw SQL against the same file can still write an overlap.
• A Clock can be told its floor. raise_floor is a required trait method, not a defaulted one, because "strictly increasing across restarts" depends on what the database already holds and no clock can see that by itself.
• CHUNK_BUDGET's 3 ms has exactly three exemptions, and they are stated with the bound rather than scattered across rustdoc: write_bulk_atomic, archive() and rebuild_current are atomic by contract and cannot be chunked without breaking the guarantee each exists to provide. Callers who need the latency bound instead of the atomicity have bulk_import.
• A narrowing predicate is not free if it changes the plan. A covering index is chosen for containing the columns, not for discriminating between rows, so the more columns one carries the more queries it silently captures. This has now bitten the same codebase three times; index-sensitive queries get a test that asserts their EXPLAIN QUERY PLAN.
• close() is how you learn the write actor died. It propagates the writer's Result and writes the final snapshot; dropping instead is legal and costs one snapshot, which is slower rather than wrong (Doctrine VI), so Drop warns and does not assert.
• Concepts are never deleted, and that guard protects more than the ledger. trg_concepts_guard_delete is unconditional (D-022), so concepts.rowid stays dense — which is the only reason the external-content FTS index survives a VACUUM. Any future concept archival or erasure has to rebuild the index (D-071).
• A 'D' row in transaction_log is corruption. Nothing writes one: Doctrine V forbids the delete and the archive moves rows rather than logging their removal. The fold refuses it rather than folding it as a tombstone (D-072).
• raw() is an escape hatch, and actor containment above it is a convention. query_only protects read_conn(); nothing protects a connection you open yourself. This is the same limit §4.2 states for the overlap guard — one fact, not two: the storage layer permits what this API refuses.

Independent of all four: the R15 upstream report against libSQL 0.9.30, outstanding longest and blocked by nothing.

Documentation

• Architecture specification — normative surfaces: §4 (schema) and Appendix A (API). One file per section.
• Implementation plan — delivery status, open items, defect register. Start at §8.5 (the 2026-07-30 review, with reproductions), §8.6 (bounds that are stated and not bounded) and §9 (the four waves).
• Python bindings — §14: the async→sync boundary, the error tree, what abi3 costs, and why Subgraph stays opaque
• Python bindings plan — P0 … P8, each with its delivery note and the corrections that phase made to the plan
• Decision register — D-001 … D-109, each with its rationale and the disqualifying flaw of the alternative

License

See LICENSE for details.