chore(follow-ups): anchor CTA cross-link + requirements upper-bound pins

Two of the post-merge follow-ups documented in
specs/004-berkshire-test/tasks.md, knocked out together:

1. Anchor essay Section 8 "Run any new proposal through the diagnostic"
now mentions both lab versions: the deterministic 5-question version
(existing link) AND the AI-powered narrative version. Wording
describes the tradeoff (fast determinism vs longer-form analysis)
so the reader picks the one that matches their use case.

2. gradio-apps/compounding-test/requirements.txt — added upper-bound
pins to all 9 dependencies. This is the same class of bug as the
Gradio 4→5 surprise that bit us during initial Space build:
gradio>=5.0,<6
transformers>=4.45,<5
torch>=2.4,<3
accelerate>=1.0,<2
huggingface_hub>=0.27,<2
anthropic>=0.39,<1
spaces>=0.30,<2
python-dotenv>=1.0,<2
pytest>=8.0,<9
When a major version of any of these lands, the Space pip-install
will fail loud on the cap rather than installing a broken combo.

Verification:
npm run build clean (anchor essay rebuilt with new link text)
pytest 63 passed, 1 skipped (unchanged)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

requirements.txt

@@ -1,22 +1,23 @@
 # Core (always required)
 # Gradio 5.x is required — earlier 4.x versions import the now-removed
 # `HfFolder` symbol from `huggingface_hub`, breaking the Space at startup
-# when paired with huggingface_hub>=1.0.
-gradio>=5.0
-python-dotenv>=1.0
-pytest>=8.0
+# when paired with huggingface_hub>=1.0. Upper bound (`<6`) prevents the
+# same silent-break class of bug when 6.0 lands.
+gradio>=5.0,<6
+python-dotenv>=1.0,<2
+pytest>=8.0,<9
 
 # Anthropic backend (only needed if MODEL_PROVIDER=anthropic)
-anthropic>=0.39
+anthropic>=0.39,<1
 
 # HuggingFace API backend (only needed if MODEL_PROVIDER=huggingface)
-huggingface_hub>=0.27
+huggingface_hub>=0.27,<2
 
 # ZeroGPU backend (only needed if MODEL_PROVIDER=zerogpu, i.e. running
 # on a HuggingFace Pro Space with on-demand A100/H100 allocation).
 # These are heavy (~2GB total via torch); local-only users who do not
 # plan to use the zerogpu backend can omit these.
-spaces>=0.30
-transformers>=4.45
-torch>=2.4
-accelerate>=1.0
+spaces>=0.30,<2
+transformers>=4.45,<5
+torch>=2.4,<3
+accelerate>=1.0,<2