perf(fsst): push down '%suffix' LIKE without decompression

[jackylee-ch]

LikeKind::parse handled prefix% and %needle% but not %suffix, so the kernel decompressed the
whole column for those — even though vortex-duckdb lowers suffix(col, 'x') to LIKE %x.

Takes the TODO's backward-scan route: a forward DFA cannot exit early, since a suffix match is only
decided at the last code. The reverse parsing that needs is local; the module doc has the argument.

• Tracking Issue: Support prefix/suffix/substring function pushdown to DuckDB. Fix contains() performance. #7734

Benchmarks

fsst_suffix against fsst_suffix_canonicalize, both in one bench binary so this reproduces from a
clean checkout. divan median of 100 samples, 8 repeats; match rates 0–100%. worst is the
canonicalize arm's fastest sample over the pushdown arm's slowest.

dataset	B/row	canonicalize	pushdown	ratio	worst
rare	44.3	26.18 µs	4.81 µs	5.4x	4.9x
log	127.8	30.84 µs	7.48 µs	4.1x	3.6x
json	96.6	19.27 µs	4.69 µs	4.1x	3.7x
cb	92.5	23.31 µs	6.52 µs	3.6x	3.5x
path	37.8	8.62 µs	5.60 µs	1.5x	1.4x
urls	35.1	8.29 µs	5.60 µs	1.5x	1.3x
email	21.9	6.96 µs	7.21 µs	0.97x	0.79x

email is a wash, and the noisiest case (6.42–8.75 µs across repeats). Its 9-byte suffix exceeds
FSST's 8-byte symbol cap, so the last symbol never settles the match and every row walks 2–3 tokens,
with only 21.9 B/row to amortise it. Deciding on one symbol would be 4.3x, so this is headroom rather
than a floor — say if you would rather gate on row length than carry it.

Tests

91 → 109. run_like now asserts the kernel returned Some before comparing booleans, since the
fallback returns the same answers: unhooking parse_suffix fails 13 tests. Three more mutations each
fail one new test.

AI assistance

Agentic AI assistance; I derived the parity argument by hand and checked the mutations fail.

[joseph-isaacs]

Do you have any benchmarks using real world or synthetic data?

[jackylee-ch]

Yes — fsst_suffix and fsst_suffix_canonicalize arms are in encodings/fsst/benches/fsst_like.rs,
both in one binary so the comparison runs from a clean checkout; table is in the description. The
forward DFA I first pushed lost on all seven datasets, so this is a backward walk from each row's end
now: six datasets are 1.3–4.9x faster taking each arm's worst sample, and email is a wash (0.97x).

Two things the rewrite also fixed: a code byte past the symbol table used to index out of bounds
instead of answering "no match", and the end-to-end cases were passing through the fallback — they now
assert the kernel pushed down.