fix(evaluation): restructure eval submodule and data for fail-safe clones

- move submodule modules/eval -> evaluation/framework
- use HTTPS URL and update=none so clones (incl. --recursive) never fail
  on the UZH-private framework repo; internal opt-in init documented
- move ground-truth data under evaluation/data/
- add evaluation/README.md describing the layout

Author: rschlaefli

.gitmodules

@@ -1,3 +1,4 @@
-[submodule "modules/eval"]
-	path = modules/eval
-	url = git@gitlab.uzh.ch:ai-infrastructure/evaluation.git
+[submodule "evaluation/framework"]
+	path = evaluation/framework
+	url = https://gitlab.uzh.ch/ai-infrastructure/evaluation.git
+	update = none

evaluation/README.md

@@ -0,0 +1,22 @@
+# Evaluation
+
+AI-quality evaluation assets for KlickerUZH.
+
+- `framework/` — git submodule pointing at the UZH-internal
+  [`ai-infrastructure/evaluation`](https://gitlab.uzh.ch/ai-infrastructure/evaluation)
+  harness (transport-neutral DeepEval runner). The repository is **private to
+  UZH**: the submodule is declared with `update = none`, so regular clones —
+  including `git clone --recursive` — skip it and nothing in the build, dev
+  stack, or CI depends on it being present.
+- `data/` — committed evaluation datasets (e.g. ground-truth QA cases under
+  `data/ground_truth/`). Synthetic examples only; no production or personal
+  data.
+
+UZH-internal users with GitLab access can materialize the framework explicitly:
+
+```bash
+git submodule update --init --checkout evaluation/framework
+```
+
+How the framework consumes the datasets in `data/` (ground-truth roots, tool
+profiles, metric suites) is still being wired up and documented.