pipeline-sql: run one SQL query across a stage as a merged table

[tarzan]

Running SQL through pipeline-cmd ('pg:psql -c ...') squeezes each app's own psql table into the appname | output layout, which is unreadable. This adds pipeline-sql <pipeline> <stage> ("<sql>" | --file=<path>), which runs one query against every app's DATABASE_URL via pg:psql -f and prints a single merged table with the app name as the first column: aligned on a terminal, appname;col1;col2 when piped, buffered and sorted by app.

psql is driven entirely through meta-commands in the script file (heroku's pg:psql exposes no psql flags): unaligned tab-separated output, no footer, ON_ERROR_STOP, and a sentinel \echo line so anything ~/.psqlrc prints first is discarded. stdout and stderr are captured separately, so psql NOTICEs never become rows and are forwarded to stderr with the app name; failed apps get their error in place of rows, zero-row apps are skipped with a count (or shown as one empty row with -a), and a header that differs between apps is reported as schema drift.

The chunked fan-out core of run_pipeline_workers is extracted into make_run_tmpdir / fan_out_workers / report_skipped so all three stage commands share it; pipeline-cmd and config-replace output is unchanged and their tests are untouched. 22 new bats tests cover the command (69 total, shellcheck clean), and it was exercised end-to-end against a real local psql through a stubbed heroku.

🤖 Generated with Claude Code