Fix generate-enums workflow: live logs, drop Sources line, camelCase protocolSpecifics (#2084)

A few related fixes to the `generate-enums` workflow and its generated
output.

### 1. Stream fetch logs live in CI
The `Fetch server data` step appeared to hang with no output, because
Python block-buffers stdout when it's a pipe (as in GitHub Actions).
Setting `PYTHONUNBUFFERED=1` forces unbuffered output so progress
streams live. (Does not make the step faster — it's intentionally serial
and rate-limited.)

### 2. Drop the `Sources:` line from generated docs
`generate_device_catalog.py` emitted an internal `Sources:
atlantic_cozytouch, somfy_europe` line into the `device-types.md` header
note. It doesn't belong in user-facing docs — removed.

### 3. camelCase `protocolSpecifics` keys
The nested keys inside `protocolSpecifics` (`clusterId`, `attributeIds`,
`commandIds`, `manufacturerId`) were snake_case in the committed JSON —
a leftover from the v1 fetch path that ran responses through
`humps.decamelize`. Every other key in the data files is camelCase.
Regenerated both `somfy_europe.json` and `atlantic_cozytouch.json` so
they're consistent (pure casing change, no data differences).

The catalog generator read the old snake_case `cluster_id`; updated to
`clusterId` so cluster annotations aren't silently dropped. With both
the data and generator on camelCase, `device-types.md` regenerates
identically to `main` — verified.

This also makes the weekly automated regeneration PR produce clean diffs
instead of thousands of casing-flip lines (the symptom seen in #2083).

### 4. Trailing newline in fetched JSON
The data files lacked a final newline, which `end-of-file-fixer`
requires — so the auto-generated PR would fail the lint workflow.
`fetch_server_data.py` now writes the trailing newline, and the
committed files were corrected.

### 5. Format generated files before opening the PR
Added a step that runs prek's auto-fixers (ruff, trailing whitespace,
end-of-file) on the generated outputs *before* `create-pull-request`.
This keeps the weekly auto-PR green instead of landing red and needing a
manual formatting fix — exactly the failure mode that #4 would otherwise
cause every week. Scoped to the generated files; type checks and the
branch guard are skipped since it's a formatting pass, not a gate.
(Note: the full lint matrix still runs separately on the opened PR via
`lint.yaml`.)

Author: iMicknl

.github/workflows/generate-enums.yml

@@ -36,6 +36,7 @@ jobs:
 
       - name: Fetch server data (Somfy Europe)
         env:
+          PYTHONUNBUFFERED: "1"
           OVERKIZ_USERNAME: ${{ secrets.OVERKIZ_USERNAME }}
           OVERKIZ_PASSWORD: ${{ secrets.OVERKIZ_PASSWORD }}
         run: uv run python utils/fetch_server_data.py --server somfy_europe
@@ -46,6 +47,22 @@ jobs:
       - name: Generate device catalog docs
         run: uv run python utils/generate_device_catalog.py
 
+      - name: Format generated files
+        # Auto-fix formatting (ruff, trailing whitespace, end-of-file) on the
+        # generated outputs so the opened PR passes the lint workflow. Type
+        # checks and the branch guard are skipped — this is a formatting pass,
+        # not a gate. prek exits non-zero when it modifies files, which is the
+        # expected outcome here, so the result is committed by the next step.
+        env:
+          SKIP: no-commit-to-branch,mypy,ty
+        run: |
+          uv run prek run --files \
+            docs/data/*.json \
+            pyoverkiz/enums/*.py \
+            docs/ui-profiles.md \
+            docs/device-types.md \
+          || true
+
       - name: Create pull request
         uses: peter-evans/create-pull-request@v8
         with: