fix: split camelCase and truncate community slugs at word boundaries

[SHudici]

What

_to_slug() lowercases its input and hard-slices at 30 characters, so a camelCase class name produces an unreadable, mid-word community name. Observed on a real ~4.8k-node graph:

services-testbuildpreanalysispromptbloc   # from TestBuildPreAnalysisPromptBlock...

Two problems compound here: camelCase survives lowercasing as one giant "word" (the existing [^a-z0-9]+ substitution never fires inside it), and the [:30] slice cuts wherever it lands.

How

_to_slug() now:

1. normalizes punctuation to spaces,
2. splits camelCase/snake_case into words via the existing _split_name() helper,
3. joins with hyphens,
4. truncates at the last complete word within the 30-char budget.

A single word longer than the budget still falls back to a hard cut (best effort).

TestBuildPreAnalysisPromptBlock  ->  test-build-pre-analysis-prompt   (was: testbuildpreanalysispromptbloc)
AuthService                      ->  auth-service                     (was: authservice)
rate_sheet                       ->  rate-sheet                       (unchanged)

Community names derived from dominant classes become hyphenated and readable; names from file stems and already-split keywords are unchanged.

Testing

• New TestToSlug class: 9 tests covering camelCase splitting, exact-boundary cuts, mid-word cuts, overlong single words, punctuation, and empty input.
• tests/test_communities.py passes (36 passed on Windows; the teardown errors are the pre-existing Windows unlink issue addressed separately in test: make the suite pass on Windows #596).

Known trade-offs (kept out of scope)

• Word-boundary truncation can collide slightly more often than a hard cut when two long names share every complete word within the budget (service_SupercalifragilisticAlpha / ...Beta both become service-supercalifragilistic). The hard cut already collided for shared 30-char prefixes; name-level dedup is handled in fix: name community splits from their members and dedupe community names #603.
• _split_name does not split consecutive-uppercase acronyms (HTTPServer → httpserver, unchanged from current behavior for every existing caller). Improving that helper would also change keyword extraction, so it belongs in its own change if wanted.