fix(generate): read WORKSPACE_DIR dynamically so relocated output isn't lost

[kevin9327]

What

After the workspace is relocated at runtime (POST /settings/paths), 3D generations kept writing their output into the previous workspace directory, while /workspace/{path} served files from the new one — so the finished mesh 404s and never appears in the app.

Why it triggers

api/routers/generation.py bound the workspace path with from services.generator_registry import WORKSPACE_DIR, capturing the Path at import time. POST /settings/paths → generator_registry.update_paths() rebinds the WORKSPACE_DIR module global (and repoints every loaded generator's outputs_dir):

# services/generator_registry.py
_self_module.WORKSPACE_DIR = workspace_dir
for gen in self._generators.values():
    gen.outputs_dir = workspace_dir

Because _run_generation used the stale imported name, it recomputed coll_dir = WORKSPACE_DIR / collection against the old directory and set gen.outputs_dir back to it — overriding the value update_paths had just written — then built output_url relative to the same stale path. /workspace/... reads the current WORKSPACE_DIR dynamically, so it can't find the file.

This affects both /generate/from-image and /workflow-runs/from-image (both funnel through _run_generation).

Fix

Reference the module attribute registry.WORKSPACE_DIR at call time instead of a name captured at import — the same way optimize.ply_to_splat (comment: "workspace dir may change at runtime") and the settings router already read it.

Verification

• Added api/tests/test_generation_router.py: drives _run_generation with a fake generator after relocating the workspace, and asserts the output is filed under the current workspace (and that output_url matches). Fails before, passes after.
• python -m unittest discover -s tests in api/ (venv with fastapi + python-multipart + httpx): all pass, 0 failures.