lanes-index never bootstraps a Postgres store: psql's "does not exist" error is read from its last line, the caret, so the store is called unreachable #244

Found during the delta check of #219. The defect is on main as well. #219 neither causes it nor fixes it.

The defect

lanes-index never bootstraps a Postgres store whose tables are missing. It calls the store unreachable and writes nothing, so sync exits 1 and export gives up with why = unreachable: followed by spaces and ^.

Line numbers are on main f5c3bcd. On #219's head 5bf29d9 they are :1321, :1355, :1365 and :989.

1. PostgresStore._run keeps only the last line of psql's stderr (lanes-index:1284). It raises StoreUnusable("unreachable", err[-1]). When the server gives a position, real psql prints three lines:

   • the ERROR: line;
   • a LINE 1: excerpt;
   • a caret line.

   So the detail that _run keeps is the caret.

2. The "does not exist" check never matches (lanes-index:1313). open_read looks for "does not exist" in exc.detail, and the detail is only ^. So a missing schema_version table is reported as unreachable instead of wiped. open_write (:1323) bootstraps only on wiped, so the DDL never runs.

3. ddl() has no CREATE SCHEMA (lanes-index:985). Even if (2) matched, open_write would fail on a schema that does not exist yet. Its first CREATE TABLE IF NOT EXISTS "<schema>".schema_version fails with ERROR: schema "<schema>" does not exist, and _run would again keep only the caret line.

The tests miss all of this because FAKE_PSQL in tests/test_lanes_index.py never fails. It models a missing store as an empty answer to the version read. That takes the "no schema_version row" path to wiped, and the fake's DDL "creates" the version.

Repro (postgres:16, 16.15)

Setup

• Server: a throwaway postgres:16 container (--network none, trust auth). psql is reached through a wrapper that runs docker exec -i <container> psql "$@".
• Configuration: url= and reader_url= both postgresql://postgres@localhost/postgres, and schema= a fresh name for each case.
• The helper: LANES_EDIT is a stub lanes-edit.sh that answers workstation (Eagle, env) and prints one TREE line for worktrees --all.
• The environment: every LANES_* and TMUX variable is unset.
• The commands: lanes-index sync --source inventory, then lanes-index export --out <dir> --worktrees.

The spaces before ^ are shortened above. The real detail is psql's caret line, indented to the error's column.

psql's own stderr for the indexer's first read, COPY (SELECT version FROM "<schema>".schema_version) TO STDOUT; (rc=3):

ERROR:  relation "repro6_absent.schema_version" does not exist
LINE 1: COPY (SELECT version FROM "repro6_absent".schema_version) TO...
                                  ^

open_write's own script, built from ddl("postgres", "repro6_absent"), against a schema that does not exist (rc=3):

ERROR:  schema "repro6_absent" does not exist
LINE 1: CREATE TABLE IF NOT EXISTS "repro6_absent".schema_version (v...
                                   ^

What a fix needs to decide

• How _run reports a failure. It could keep the ERROR: line rather than the last line: the first line that starts ERROR: or FATAL:, or the whole stderr. It could also run psql with -v VERBOSITY=terse, which prints the error on one line. Then a missing table classifies as wiped, and other failures name their real cause instead of a caret.
• Whether open_write creates the schema (CREATE SCHEMA IF NOT EXISTS) or refuses with a message that names it. The writer role may not have CREATE on the database, so an explicit refusal may be the right answer where QA provisions schemas.
• A test the fake cannot fake. A fake psql that fails as real psql does (rc=3 and the three-line error) for a missing relation, so the wiped path and the bootstrap are proven against psql's real shape.