Use chunked numeric posting buffers

[platypii]

Summary

• replace per-token numeric Maps and JavaScript posting arrays with append-only typed-array chunks
• group postings by token suffix with a stable counting sort while draining each prefix
• reuse counting-sort scratch storage and release chunks as prefixes are written
• retain the existing string-Map path for custom tokenizer configurations

WildChat benchmark

Same 15,587,491,031-byte source, 3,199,860 rows, default index settings, and measurement command on both commits.

Metric	master (6f45d8b)	this branch (c22ccfa)	Change
Total build time	130.01 min	36.75 min	71.7% lower / 3.54x faster
Read, tokenize, accumulate postings	~76.36 min	~20.15 min	~73.6% lower / 3.79x faster
Group, sort, encode, and write	~53.65 min	~16.61 min	~69.0% lower / 3.23x faster
Peak RSS	40.01 GB	21.61 GB	46.0% lower
Output size	3,622,096,587 bytes	3,622,096,587 bytes	unchanged

The phase boundary is approximate because the benchmark records one-minute heartbeats rather than explicit internal phase timers. Total wall time and peak RSS come directly from performance timing and macOS /usr/bin/time.

The branch output is byte-for-byte identical to the master artifact.

Test plan

• npm test (153 tests pass)
• npm run lint
• full WildChat index build
• cmp against the master-generated index artifact