T-298: perf(quartet_concordance): hoist buffer resize outside split loop

[ms609]

Summary

• Hoist n0/n1 buffer resize outside the split loop: was resize-per-char-per-split, now resize only when a new character has more states than current capacity (once per character in the worst case).
• Compute max_state in the existing char-column caching pass — no extra loop added.
• Add dev/benchmarks/bench_quartet_concordance.R for timing and profvis flame-graph.

Analysis notes (T-298)

Question	Finding
Flat array vs NumericMatrix for output	No difference: output writes are O(n_splits × n_chars), outside the inner loop; -O2 inlines Rcpp accessor to a single pointer dereference
int vs double	n0/n1 are already int; output correctly uses double (decisive values can be half-integers)
Further micro-opts (split col ptr cache)	Compiler hoists base + nrow * s under -O2; no profile evidence of residual overhead — left alone

Benchmark results (Windows, R 4.5.2, seed 42)

Size	Taxa	Splits	Chars	Time/call
small	25	23	50	<1 ms
medium	100	98	200	14 ms
large	300	298	400	186 ms

Test plan

• GHA agent-check passes (run 25777319791)
• No regression in test-QuartetConcordance tests

🤖 Generated with Claude Code