Cascade Utf8 dict codes so they bitpack (#303)

[dfa1]

What

The per-chunk Utf8 dictionary path emitted its codes as a raw vortex.primitive leaf (full U8/U16/U32 width per row), while the Rust reference bitpacks them. This was the largest remaining per-column gap after #299 (the street columns in #303).

Encoding tree for a street column (Cross Street 1), before vs after:

vortex-jni                     vortex-java before             vortex-java after
  DICT                           per-chunk vortex.dict          per-chunk vortex.dict
    codes: fastlanes.bitpacked     codes: vortex.primitive  ❌     codes: fastlanes.bitpacked  ✅
    values: vortex.fsst            values: vortex.fsst            values: vortex.fsst

How

encodeUtf8Cascade exposed the values pool as an open child (#299 Finding 2b) but left codes as a raw leaf. Now both children are open: codes (index 0) as a typed primitive array the cascade bitpacks, and the values pool (index 1) for the FSST/VarBin competition. The primitive dict path already did this; only the Utf8 path was missing it.

No wire/reader change — the reader already decodes bitpacked dict codes (it reads Rust files that use them; conformance + interop suites cover it).

Impact (real nyc-311 corpus, off main)

	vortex-java	vs vortex-jni (1762 MB)
Before (main)	1934.23 MB	1.10×
After	1862.23 MB	1.06×

−72 MB — larger than the ~54 MB street-column estimate because it helps every per-chunk Utf8 dict column. Composes with #302 (that fix bitpacks the FSST offset/length children — different columns); together they put vortex-java well under the reference.

Testing

• Updated unit test: encodeCascade now exposes two open children (codes + values).
• Dict round-trip + 217 Java-writes-Rust-reads interop tests pass.
• Full ./mvnw verify green.

Addresses #303 (finding #1 — bitpacked codes). Finding #2 (per-chunk vs global dict fallback for these columns) remains open there as separate upside.

🤖 Generated with Claude Code