db-engine-selection (M0) — recommend, don't score

M0 runs at design / /claude-db:start time, when there may be no schema yet — only a described workload. It answers "which database should this be?" It is not a scored module: it emits no findings, has no axis, no severity, and never enters score.mjs or either of the two scores. Its output is a separate recommendation contract (below). It walks the decision tree in references/engine-selection-tree.md.

Inputs it gathers

From the user's description (or detected stack via references/detection-signals.md):

• Access patterns — point lookups by key, range scans, complex multi-entity joins, full-text / semantic search, time-ordered analytics, graph traversal, fan-out reads/writes.
• Consistency & integrity needs — strong/transactional vs eventual; multi-row invariants; need for DB-enforced referential integrity.
• Scale & shape — expected size, write rate, read/write ratio, cardinality, whether data is append-only/time-series, vector/embedding workloads.
• Operability & team — managed vs self-hosted, serverless/edge, existing expertise, platform constraints (cross-checks db-platform-fit M21).

How it decides (tree summary; full tree in the reference)

1. Relational by default for transactional, multi-entity, integrity-heavy workloads (Postgres as the safe default; MySQL/MariaDB where the ecosystem dictates).
2. Document (Mongo/Firestore) when the data is aggregate-oriented, read by one access pattern, and embedding beats joining — and the team accepts app-enforced integrity.
3. Key-value (Redis/DynamoDB) for known-key point access, caching, sessions, high-throughput simple ops.
4. Wide-column (Cassandra/Scylla) for massive write-heavy, table-per-query, partition-first workloads.
5. Vector (pgvector/Qdrant/etc.) for semantic search / RAG — often alongside a primary store, not instead of one.
6. Time-series (Timescale/ClickHouse/Influx) for append-only metrics/events with time-range analytics.
7. Graph (Neo4j) when traversal depth/relationship queries are the core workload, not an afterthought. Polyglot is a valid answer: name each store and its job. Prefer "Postgres + an extension" (pgvector, JSONB, partitioning, FTS) before adding a second engine, when one engine credibly covers the workload.

Recommendation contract (what M0 emits — NOT a finding)

A structured recommendation object, rendered by /claude-db:start, with fields:

• recommended — { paradigm, engine, platform? }.
• rationale — which access patterns / consistency / scale facts drove it, tied to the tree branch.
• runner_up — { paradigm, engine } and the explicit trade-off of choosing the recommendation over it (what you give up).
• assumptions — the workload facts assumed; if a fact is unknown, it is listed here, not invented.
• confidence — established | directional | speculative (a description-only input is at most directional); speculative when key facts are missing.
• caveats — lock-in, operability, and platform notes (defers exact prices/limits to M21 / vendor page). This object is rendered as guidance; it does not conform to schema/finding.schema.json and is absent from audit-report.schema.json.

Verification / honesty

• Never fabricate benchmarks, latency, throughput, prices, or row-count ceilings to justify a pick. Compare on documented properties (consistency model, FK support, index types, operability, lock-in) and band any magnitude high|medium|low. Cite a vendor page via WebFetch only by quoting it.
• Always name the runner-up and the trade-off — a recommendation without an alternative is a red flag.
• If the workload is under-specified, say so and ask (or mark confidence: speculative); do not force a pick from thin air.
• M0 produces no score and never caps anything — it is advisory by construction. When an existing schema is present, defer correctness judgments to the scored modules (M1..M22) and the auditors.