Obfuscate EXTRACT field keywords

[joelmarcotte]

Summary

• Customer hit duplicate DBM query signatures because EXTRACT(epoch FROM ...) was obfuscated differently depending on capture path: pg_stat_statements normalizes to EXTRACT($1 FROM ...) (replaced with ?), while pg_stat_activity preserved epoch as an unquoted identifier.
• The obfuscator now tracks EXTRACT( context and rewrites the field argument to ? when it matches a known PostgreSQL field name (epoch, year, month, dow, isodow, microseconds, timezone_hour, etc.), so both code paths converge on EXTRACT(? FROM x).
• Conservative match: only known field keywords are replaced. Bare epoch outside an EXTRACT call (e.g. a column reference) and unrecognized field names are left alone, so user identifiers aren't accidentally rewritten.

Test plan

• go test ./... passes (existing + new)
• New unit cases in obfuscator_test.go cover the standalone Obfuscate path: positive cases (lowercase, uppercase, mixed-case fields), positional parameter regression, and negative cases (epoch as a column, unknown field name)
• New JSON fixtures in testdata/postgresql/select/ cover ObfuscateAndNormalize:
  • extract-epoch-field.json — primary regression test for the customer issue
  • extract-various-fields.json — year/month/dow/microseconds/timezone_hour in one query
  • extract-positional-parameter.json — pg_stat_statements form
  • extract-quoted-string-field.json — EXTRACT('epoch' FROM x) (string literal alternative)
  • date-trunc-string-literal.json, date-part-string-literal.json — already obfuscated, locked in
  • extract-field-name-as-column.json — confirms epoch/year outside EXTRACT remain untouched
• BenchmarkObfuscator* shows no measurable regression (the change adds two string compares per non-whitespace token)

🤖 Generated with Claude Code