Fix 500 on package retry/upgrade due to duplicate environment_packages rows

[Copilot]

POST /environments/{id}/packages/{name}/retry crashed with sqlalchemy.exc.MultipleResultsFound whenever a package had been installed more than once — each failed install silently created a new DB row, and scalar_one_or_none() exploded on the duplicates.

Root cause

No uniqueness constraint on (environment_id, package_name). Every call to add_package() unconditionally inserted a new row, so a failed install followed by a re-install produced duplicates. retry_package_install, upgrade_package, and the background task workers all used scalar_one_or_none() which raises on multiple results.

Changes

Prevent duplicates at source

• add_package() now upserts: if a row for (env_id, package_name) already exists it resets install_status → pending, clears install_error/installed_version, and returns the existing record instead of inserting a new one.

DB constraint

• Added UniqueConstraint("environment_id", "package_name") to EnvironmentPackage model.
• New Alembic migration (f1a2b3c4d5e6) deduplicates any existing rows (keeping the highest-id record per pair) then applies the constraint.

Fix scalar_one_or_none() callsites

• retry_package_install, upgrade_package (router) and _install_package_inner, _upgrade_package_inner (tasks) all changed to .scalars().first() — defensive against pre-migration data.
• remove_package likewise uses .scalars().all() for the same reason.

Fix missing commit before dispatch

• install_package and upgrade_package now call db.commit() before dispatch_task(), per the CLAUDE.md critical pattern (background thread uses a separate session and won't see uncommitted data).

Tests

• test_install_package_twice_reuses_existing_record — verifies second install reuses the same row, not a new one.
• test_retry_after_failed_install_does_not_500 — verifies retry returns 200 after a failed install.

[raffelino]

lgtm