* fix: add id-token write permission and fix invalid inputs in GC workflow
The claude-code-action@v1 uses OIDC for OAuth token auth, which requires
id-token: write in workflow permissions. Without it ACTIONS_ID_TOKEN_REQUEST_URL
is not set and the action fails after 3 retry attempts.
Also moves model and allowed_tools from unsupported top-level inputs into
claude_args where they are valid CLI flags for claude-code-action@v1.
* fix(gc): show agent output and explicit dry-run issue posting
- Add show_full_output: true so Claude responses appear in CI logs
- Replace vague dry-run instruction with explicit gh issue create/comment
commands so output has a concrete destination (tracking issue)
- Bake DRY_RUN value into prompt at workflow-expansion time to eliminate
ambiguity about env var reading
* docs: mark cloud-auth-architecture.md as superseded by ADR-003
* fix(docker): guard community seeder cd and add ENCRYPTION_MASTER_KEY default
- Wrap `cd $SEEDERS_DIR/community` in an existence check so the
data-engine container no longer crashes when local-seeders/community/
is an empty or uncloned nested git repo. The container was entering
Docker's exponential-backoff restart loop (2s→60s) on every boot.
- Add ENCRYPTION_MASTER_KEY to the wwv environment block with a
local-dev fallback, matching the AUTH_SECRET pattern. Without it the
Next.js instrumentation hook throws at startup and the app never
serves a request.