fix: make segment joins seam-aware

[bukinoshita]

Summary by cubic

Make segment joins seam-aware to prevent illegal consonant clusters, vowel pileups, and bad final letters. Inserts tiny linking phonemes and updating finals; sampled outputs show illegal starts/ends and 3+ vowel runs drop to ~0 with a small uniqueness gain.

• Bug Fixes
  • Validate C+C seams against whitelists and insert a linking vowel when illegal (INITIAL_CLUSTERS, MEDIAL_CLUSTERS).
  • For V+V seams, block 3+ runs and aa/ii/uu by inserting l/n/r.
  • Special-case qu so it is followed by exactly one vowel; repair with vowel or vowel+consonant bridges.
  • Remove illegal finals by updating CODAS/RIMES (no word ends in j, q, w, h, v).
  • Replace raw concatenation in buildPattern with joinSegments; add can-join and cluster utilities with tests.

^{Written for commit 2b738d6. Summary will update on new commits.}

[cubic-dev-ai]

cubic analysis

No issues found across 11 files

Linked issue analysis

Linked issue: BU-2307: 2. Make segment joins seam-aware (no illegal clusters or vowel pileups across seams)

Status	Acceptance criteria	Notes
✅	Validate C+C seams against initial/medial whitelists (disallow illegal consonant clusters at seams).	PR adds INITIAL_CLUSTERS and MEDIAL_CLUSTERS, canJoin uses them, and tests cover both unit cases and sampled words to ensure no illegal initial/medial clusters.
✅	Prevent vowel+vowel seams that create runs of 3+ vowels or banned vowel pairs (aa/ii/uu).	canJoin enforces vowel-run constraints and banned pairs; joinSegments inserts linking consonants; unit tests verify both canJoin and join behavior and sampled words show none of these vowel errors.
✅	Special-case qu so it is followed by exactly one vowel (repair when needed).	canJoin checks left.endsWith('qu') and joinSegments provides QU_LINKING_BRIDGES/QU_LINKING_VOWELS; tests exercise qu+consonant and qu+vowel cases.
✅	Remove or restrict illegal finals so no word ends in j, q, w, h, or v.	phoneme arrays were updated to drop illegal finals and tests assert CODAS/RIMES/TAILS contain no entries ending in those letters; sampled words tests assert no outputs end with those letters.
✅	Replace raw concatenation in buildPattern with joinSegments and add can-join/cluster utilities and tests.	buildPattern now delegates to joinSegments; new utilities (can-join, clusters) and many unit tests were added, and joinSegments behavior is covered by deterministic and behavioral tests.
⚠️	Acceptance criterion: illegal word-start and word-end cluster rates drop to ~0 on a 200k-word sample.	The PR adds thorough unit tests and sampled-word assertions that show zero offenders on the provided samples, but the tests/sample sizes are much smaller than the 200k-word sample specified in the issue (pattern.spec samples ~500 per pattern; generator tests sample up to 5k). The claimed large-sample measurement is not shown in the repo tests here.

_{Re-trigger cubic}