gdrive connector with google picker api

[martyna-mindsdb]

Linear ticket: https://linear.app/mindsdb/issue/ENG-687/google-drive-google-picker-api

---

Demo video: https://www.loom.com/share/6c689e3708664779aa13900ae20e4e49

---

Adds the server-side infrastructure for the Google Drive Picker, including endpoints to persist, query, and untag picked-file grants on a connection record, per-project scoping so the agent only sees files relevant to the current project, and delivery of the Google Picker API key required by the desktop app.

This PR also includes several reliability improvements from code review: a shared per-connection lock to eliminate read-modify-write races, preservation of picked files and custom connection labels across OAuth reconnects and token refreshes, and consolidation of duplicated connection-loading logic.

File	What & why
cowork/schemas/connectors.py	Added PickedFile (including project tags) and PatchPickedFilesBody request models.
cowork/common/settings/app_settings.py	Added the google_picker_api_key (GOOGLE_PICKER_API_KEY) setting used by the browser-based Google Picker widget.
cowork/services/connectors/oauth/config.py	Added a uses_picker flag to GoogleServiceConfig so the credentials endpoint determines whether to include the Picker API key from configuration instead of relying on a hardcoded engine-name check.
cowork/api/v1/endpoints/connectors/oauth.py	Updated the /credentials response to include picker_api_key when uses_picker is enabled.
cowork/api/v1/endpoints/connectors/connections.py	Added PATCH .../picked-files to merge Picker selections and DELETE .../picked-files/{file_id}?project=... to remove a file's association with a single project.
cowork/services/connectors/connections.py	Implemented the core picked-file logic. merge_picked_files unions project tags instead of overwriting them, while remove_picked_file only removes a project association and never revokes the underlying grant. Added a shared per-(engine, name) lock around all read-modify-write operations (merge_picked_files, remove_picked_file, patch_token) to eliminate concurrent update races. Extracted shared _read_record and _load_picked_files helpers to remove duplicated logic, and renamed the vault field to _picked_files so it is treated as an internal field rather than an agent-visible credential.
cowork/services/connectors/oauth/google.py	Preserved _picked_files and _label across OAuth reconnects and token refreshes instead of overwriting them. Wrapped updates in the shared connection lock to prevent races with concurrent Picker operations.
cowork/services/connectors/persist.py	Updated persist_connection to preserve _picked_files during full-record saves, wrapped it in the shared lock, and also synchronized set_connection_label for consistency.
cowork/services/connectors/vault_lock.py	Added a shared per-(engine, name) lock registry used by all connection write paths to serialize concurrent operations. Includes cleanup when connections are deleted to prevent the registry from growing indefinitely.
cowork/harnesses/anton_harness/harness.py	Injected project-scoped picked-file guidance into the agent's system prompt. Files are only exposed when tagged to the current project, while connection-wide files remain visible everywhere. Also merged two vault-scan loops into one to reduce directory I/O on every chat turn.
uv.lock	Lockfile bump.

[ZoranPandovski]

Move the Google Drive logic out of the harness and into the connector service. (the harness parses _picked_files JSON, checks engine != "google_drive", and applies project-scoping rules itself. This is connector logic, not agent logic

[martyna-mindsdb]

Thanks. I moved this from harness to connections and tested the entire flow successfully. Done.