Add turbopuffer sink documentation

[nstewart]

Motivation

This adds comprehensive documentation for sinking Materialize results to turbopuffer, a vector database service. The documentation covers the complete pipeline: using a Kafka sink to export data from Materialize, then using the mz-tpuf-sink Python library to apply those changes to turbopuffer namespaces.

Description

This PR adds a new documentation page (doc/user/content/serve-results/sink/turbopuffer.md) that guides users through:

1. Setup: Creating Kafka and schema registry connections, building a materialized view as the search document, and configuring a Kafka sink with Debezium envelope for change capture.

2. Running the sink: Installing and configuring mz-tpuf-sink, a Python library that consumes from Kafka and applies changes to turbopuffer with transaction atomicity guarantees.

3. Adding embeddings: Using transforms to compute embeddings (or other derived attributes) only when relevant columns change, avoiding unnecessary recomputation.

4. Validation: Querying the namespace and testing deletes to confirm the pipeline works end-to-end.

5. Operational details: Covering transaction atomicity, at-least-once delivery with idempotent writes, column-level updates, and type mapping between Materialize and turbopuffer.

The documentation also updates doc/user/data/sink_external_systems.yml to list turbopuffer as a supported external system alongside other sinks like Kafka and OpenSearch.

Key design points documented:

• The sink applies each Materialize transaction atomically to turbopuffer
• Transforms recompute embeddings only when their source columns change
• The process requires a Materialize connection to track write frontiers across partitions
• Exactly one process should run per topic/namespace pair to maintain transaction boundaries
• Column-level patches preserve attributes written from outside the pipeline

Verification

This is documentation-only. No code changes or tests are required. The documentation follows existing patterns in the Materialize docs and references the actual mz-tpuf-sink library on GitHub.

https://claude.ai/code/session_011Anxe7er6ZK6P62jjWqyfG

[nstewart]

Reviewing the sink source to write the correctness section surfaced two discrepancies in mz-turbopuffer-sink itself. Both are filed there, and neither blocks this PR:

• MaterializeInc/mz-turbopuffer-sink#1 — to_attr has no uuid.UUID branch, so a non-key uuid column reaches the writer as a Python object against an attribute declared string.
• MaterializeInc/mz-turbopuffer-sink#2 — schema.py's time-millis/time-micros mapping is unreachable. mz_avro has no time-of-day SchemaPiece, so the logical type is dropped when the schema is round-tripped for registration, and Materialize publishes time as an untagged long.

The type mapping table in this guide documents the observed behavior for both: time as int (microseconds since midnight), matching the sink's README and its end-to-end test, and uuid as string. Neither issue changes what the table should say, so no docs change is pending on them.

---

Generated by Claude Code

[nstewart]

Closing in favor of #38167