Support reading metrics from a SQL query (DuckDB)

[msto]

MetricReader is text-IO-first today: delimited rows in, validated metrics out. For metrics that already live in a DB or in Parquet/Arrow, going through a TSV intermediary is awkward. DuckDB queries any of CSV/Parquet/Arrow/JSON/SQLite/Postgres/S3-hosted files and yields rows as dicts — exactly the shape Metric.model_validate wants.

Proposal: add a MetricReader.from_sql(metric_class, query, *, connection) classmethod (DuckDB connection or relation) that runs the query, iterates rows as dict[str, Any], and validates each through the supplied Metric subclass.

Open questions:

• Scope: DuckDB-only (cleanest row→dict story) vs a generic DB-API 2.0 cursor adapter (column-name resolution becomes the caller's problem).
• Dependency shape: optional extra (fgmetric[duckdb]), matching the remote-files pattern.
• Name: from_sql, from_duckdb, from_query?
• Headerless mode + fieldnames doesn't apply here — SQL columns are already named. Worth calling out that the SQL path is header-only by construction.

[msto]

This comment was written and posted by Claude Code at @msto's request, flagging interactions with #54 (writer append mode) so they don't get lost.

Two notes for whoever picks this up:

1. Multi-member compressed streams. Appending through xopen to .gz/.bz2/.xz produces concatenated compression members. Python's stdlib decompressors read multi-member streams transparently (feat: append mode for MetricWriter.open #54's round-trip tests cover this), but DuckDB uses its own decompression — worth verifying read_csv over a multi-member .gz before promising the "append compressed shards, then query them with DuckDB" workflow.

2. Single-header invariant. Append mode writes the header exactly once across write→append cycles and validates it on re-open, so accumulated files stay directly ingestible by DuckDB's read_csv. And if a SQL write path ever follows this issue, feat: append mode for MetricWriter.open #54's contract maps one-to-one: mode="w" ≈ CREATE OR REPLACE TABLE, mode="a" ≈ append-or-create — create the table from the metric class's fields if absent, validate columns and INSERT if present, raise on mismatch.