feat: unblock agent workflow — pager-free help + subtype narrowing

[prantogg]

Why

From an agent session (Copilot, terminal shell, local install):

pip show wkls OR python3 -c "import wkls; help(wkls)" | head -50
# … pager hangs, agent sends 'q', then echo 'test', finally:
PAGER=cat MANPAGER=cat python3 -c "..."

Agent wasted ~8 commands escaping the terminal pager that help(wkls) opens. Then spent another ~10 commands on inspect.signature + dir(Wkl) to reverse-engineer the API. Finally ran .search('san diego').wkt(), hit AmbiguousLocationError, and jumped to by_id('<uuid>') from the error message — never discovering that .search('san diego').cities().wkt() is a faster path.

What

Three changes unblock the agent workflow without rewriting anything:

1. wkls.help() — prints the module docstring to stdout. Same content as help(wkls) but no pager. Listed in dir(wkls) so agents find it. Intentionally omitted from __all__ to avoid polluting from wkls import *.
2. AmbiguousLocationError suggests subtype narrowers when candidates span multiple subtypes. E.g. for San Diego (city + county), the error now emits .cities() / .counties() alongside by_id(...). Suppressed when narrowing wouldn't reduce the set (e.g. all 18 Franklin townships in PA are localities — .cities() wouldn't help, so it isn't suggested).
3. Multi-row result-mode dir() advertises the listing methods (cities, counties, regions, countries, dependencies, subtypes). Previously only inspection verbs (count, head, etc.) showed up — so agents never saw the subtype-filter path unless they read the docstring.

Module docstring updated to lead with wkls.help() and list subtype narrowing as option (a) under "Resolving ambiguity".

Also includes a small unrelated build fix: broader .venv-* ignore + explicit hatch sdist excludes so uv build doesn't pick up local virtualenvs or test_nb.ipynb.

Test plan

• uv run pytest — 151 passed, 2 xfailed (added 3 tests: pager-free help, subtype suggestion fires, subtype suggestion suppressed when redundant)
• Smoke-tested the suggested narrower end-to-end: wkls.us.ca.search('san diego').cities().wkt() returns a valid WKT
• dir(wkls) includes help; dir(wkls.us.ca.search('san diego')) includes cities, counties, etc.
• uv build succeeds with a 35-file sdist (no .venv-* / notebooks)

Follow-ups (not in this PR)

Scoped out to keep this PR focused on the agent-trap we set out to fix. Tracked from a Copilot usability review + an ecosystem survey in docs/investigations/llm-friendly-library-patterns.md:

• .head(n) / .limit(n) should return a Wkl, not a raw sedona DataFrame. Agents naturally chain .head(3).to_dicts() and it fails because .to_dicts() is a Wkl method, not a DataFrame method. Fix is ~10 LOC — wrap the sliced DataFrame in a result-mode Wkl.
• Did you mean '<X>'? on unknown chain attributes. wkls.us.ca.sanfransisco should suggest sanfrancisco via difflib.get_close_matches. Python 3.10+ ships this for NameError/AttributeError; we'd mirror that convention for chain misses. ~20 LOC.
• Distinguishing labels for identical-looking ambiguity candidates. Two Yorks in York County, PA currently render as two identical lines in the error — append last-4-of-UUID (or similar) so visually distinct, even if the ultimate picker is still by_id.
• Ship docs/llms.txt — short index file following the Answer.AI convention (adopted by Pydantic AI, Anthropic docs, Stripe, Cloudflare). Pointers to README + per-feature markdown sections, not inline docs. Near-zero maintenance.