Specializer wave B: regexp family — corpus 484 -> 505 (TASK-53)

[claude-agent-ahrzb]

What

The regexp family lands: regexp_matches / regexp_full_match / regexp_extract / regexp_replace, the ~ / !~ operators, SIMILAR TO on values, and the wave-5-deferred star forms (* SIMILAR TO, COLUMNS('re')). Corpus goes 484 → 505 of 678 with zero wrong answers and zero replay FAILs.

The engine decision

DuckDB uses RE2; we use the Rust regex crate (new dependency) — but only behind a measured translation layer. The pins fleet ran a 98-entry differential battery (DuckDB and rust-regex side by side, identical inputs) and found:

• The whole Unicode gap is the Perl classes: RE2's \d\w\s\b are ASCII, rust's are Unicode. Fixed by a bind-time rewrite (\d→(?-u:\d) etc., in-class variants included), each verified byte-for-byte.
• The "obvious" global fix — unicode(false) — was measured to break (?i) folding parity (KELVIN sign, sharp-s). Never used.
• A bind-time reject list for the irreconcilables: \B (unservable — DuckDB itself dies at runtime on non-ASCII), (?<name>) groups, duplicate group names, repetition bounds > 1000, stacked quantifiers (a*+ — rust silently reinterprets, a wrong-answer risk, not an error-shape difference), \u, \Q\E.
• Replacement templates translate (\N→${N}, $→$$) with RE2's bizarre invalid-rewrite quirks resolved at bind time: out-of-range backref = silent full no-op; global bad escape = consume-each-match-emit-prefix.

With that recipe, all 98 battery entries are byte-identical or identically-rejected.

Measured semantics worth calling out

• ~ is FULL match in DuckDB, not the Postgres search — the binder error literally names regexp_full_match. Implementing from Postgres intuition would silently flip results.
• SIMILAR TO translates NO wildcards: 'hello' SIMILAR TO 'h%o' is FALSE (% is a literal), 'h.llo' is TRUE (regex dot live). It is sugar for regexp_full_match on the raw pattern.
• regexp_extract returns '' on no-match — never NULL — and its group index is a flat 0..9 check unrelated to the pattern's group count.
• regexp_replace's NULL-options argument makes the result NULL, while the other functions error on NULL options — a pinned asymmetry.
• Star * SIMILAR TO (unanchored name search) and * NOT SIMILAR TO (NOT full-match) are not complements — implemented as independent predicates per the pins.

Infrastructure

Program grows a prepare-time regex table (round-trips through the IR text format; the verifier checks indices and rewrite presence). ReMatch/ReExtract/ReReplace run on both backends — interp closures own their compiled Regex, cranelift reads a compiled table through the row context. Constant patterns compile at prepare, matching DuckDB's pinned bind-time error eagerness; column patterns stay clean-unsupported.

Verification

• Pins-first: 6-agent fleet, spec + JSONs committed before implementation (docs/superpowers/specs/2026-07-27-waveB-regexp-pins.md)
• 154 Rust + 557 py tests green; interp backend identical minus the deliberate backend-identity guards
• Corpus: 505 match / 171 clean-unsupported / 2 known-divergent, zero FAILs
• clippy clean on wave files; serving-bench parity gate passed, spec in the usual 1.4–2.0× band

🤖 Generated with Claude Code