GeezerVison

LGTM. This PR only changes docs, so the test gates don't apply. What I checked:

• All three new anchors resolve on the Kafka sink page in the merged tree: #delivery-retries-and-backpressure, #exactly-once-delivery-and-recovery, #output-metrics.
• The new metrics paragraph is the same in metrics.md.in and metrics.md, and it sits outside the {{output_connector_}} block, so regenerating metrics.md will keep it.
• The claim that Kafka is the only exactly-once sink holds in the code. In crates/connectors, is_fault_tolerant() returns true only for kafka/ft and the null sink. Snowflake and Confluent JDBC count too, because they run on the Kafka output connector.
• pre-commit run --files <4 changed files> passed.

LGTM (docs only, stacked on #7682).

What I checked

• Every new link target exists: kafka#delivery-retries-and-backpressure, #exactly-once-delivery-and-recovery, #output-metrics, and delta#metrics.
• The claim "today only the Kafka output connector supports exactly-once output" holds in the code. is_fault_tolerant() returns true only for kafka/ft/output.rs, plus the null sink, which drops everything anyway. Delta, Postgres, NATS, Redis, S2, DynamoDB, file and HTTP all return false.
• The metrics.md.in and metrics.md edits are identical, so regenerating with metrics.py keeps them in sync.
• pre-commit run --files <4 changed files> passed.

Commit history (medium): the second commit, "Scope the exactly-once statement…", says it is review feedback. Please squash it into the first commit before merge, to keep main's history linear and clean.