gemdb refuses to start a database without Python support #115

Related: #105 covers the editor's Shell and Run File repairing a running database. This issue is the CLI half.

Problem. The gemdb wrapper (cli.ts) and the Shell's ensureDatabase (cliMain.ts) start the stone with no Grail check. That is how a "running but no Python" database arises; PR #103 repairs it, but only at the next activation. What users see from the wrapper (measured):

• Account present, no Grail: gemdb -c … and gemdb file.py dump the whole ~150-line run block of gemdb-run.tpz as "GemStone Smalltalk Compiler Errors", ending in [1031] undefined symbol, and exit 1.
• CLI-first, before any editor bring-up: the gemdb account doesn't exist yet, so the login fails with GemStone error 4051.
• The Shell is already fine: it says "Python support is not installed in this database…". The wrapper is the part to fix.

Who hits it. Brain Freeze's tutorial has readers set up GemDB, then makes gemdb tools/seed.py the first command. A reader who runs it right after "GemDB is ready", before auto-start has filed Grail in, or after pressing Stop, gets the dump. Brain Freeze's own CI works around exactly this race (ci/setup-gemdb/run.js waits for the editor and probes with gslist, not gemdb).

Decided (2026-10-06): refuse, don't install.

• When Grail was never filed in, the wrapper and the Shell refuse to start the stone. They print e.g. "Python support is not installed in the database yet: open VS Code and press Start", and exit non-zero.
• A stone that's already running without Grail gets the same message instead of the compile-error dump.
• The check is "never filed in": grailInstalled(), which tests the stamp <root>/grail/.gemdb-grail-stamp. It is not grailFileInDue/grailNeedsUpdate, because an older Grail still runs Python.
• The bash wrapper can't call TS, so it tests the stamp file directly. Keep that path in one place: Grail and MCP versioning: gaps the new install mechanism must not inherit #110 replaces the stamp, and the swap should be one line on each side.
• External databases skip the check (the wrapper already has an external branch). The local stamp says nothing about an administrator's database.
• Rejected: the wrapper filing Grail in itself. That would mean a bash copy of the installer, with no progress UI, racing an editor.

Done when: the wrapper's -c, file and -m modes, and the Shell, each print the message and exit non-zero on a database with no stamp. This needs a test with a pty, or the wrapper's own bash tests.