feat: showcase typed spans in Streamlit demo

[w3lld1]

Summary

• replace the demo's developer-local checkpoint with a fresh-clone-compatible v2 encoder cascade
• add a cached sidebar selector for typed and binary models
• color typed spans by taxonomy category and show category, subcategory, and confidence in tooltips
• add a category legend, per-span table, and focused renderer tests for typed, untyped, and multiple spans

Related issue

Fixes #57

Type of change

• Bug fix
• Feature
• Documentation
• Tests
• Refactor or maintenance

Testing

• python3 -m ruff format --check demo/streamlit_demo.py tests/test_demo_pytest.py
• python3 -m ruff check demo/streamlit_demo.py tests/test_demo_pytest.py
• python3 -m pytest tests/test_demo_pytest.py -v (3 passed)
• python3 -m py_compile demo/streamlit_demo.py tests/test_demo_pytest.py
• git diff --check
• pytest tests/test_inference_pytest.py -v -k "not TestAnswerStartToken" — collection was attempted but this runner does not have torch installed
• Interactive model-download smoke test — not run locally because the runner lacks Streamlit/model dependencies and has insufficient disk for the model stack

Checklist

• I kept the PR focused on one change.
• I added or updated tests/docs when needed.
• I checked that no secrets, API keys, or credentials are included.

Rights & sign-off (required)

• I certify that I have the right to submit this code and that it may be
  distributed under the repository's MIT license
  (see CONTRIBUTING).

[adaamko]

Thanks — this is a careful implementation and it hits every item in #57: the fresh-clone fix, the cache-keyed model selector, per-category colors with subcategory and confidence tooltips, the legend, the per-span table, and a pure renderer with focused tests (including the right-to-left insertion-order check). I verified the full suite passes merged with current main (88 passed via both python -m pytest and python tests/run_pytest.py) and ruff is clean.

One fix needed before merge, and it's on us, not you — the issue text gave you a wrong model id: the binary v1 model on the Hub is KRLabsOrg/lettucedect-base-modernbert-en-v1 (the same lettucedect slug spelling as the v2 repos; the lettucedetect- form 404s). As written, picking "Binary spans (ModernBERT)" in the sidebar crashes on download. Please swap that one string in MODEL_OPTIONS — the issue text and docs are being fixed separately (#82).

Also please update the branch (or rebase on main): CI was recently switched to run the full test suite, and your green run predates that, so tests/test_demo_pytest.py hasn't actually been exercised by CI yet. It passes locally against current main, so this should be a formality.

Non-blocking nit: use_container_width=True on st.dataframe is deprecated in recent Streamlit (width="stretch" is the replacement); fine to leave.

[w3lld1]

Thanks — fixed the binary v1 model ID to KRLabsOrg/lettucedect-base-modernbert-en-v1 and rebased the branch onto current main.

Fresh validation on 95616f8:

• python -m pytest — 142 passed
• python tests/run_pytest.py — 142 passed
• ruff format --check demo/streamlit_demo.py tests/test_demo_pytest.py
• ruff check demo/streamlit_demo.py tests/test_demo_pytest.py
• git diff --check

[adaamko]

Model id fixed and the rebase means CI now runs the demo tests under the full-suite workflow — all green. Merging; thanks for the quick turnaround and for a genuinely well-built demo upgrade.