Over-redaction splits a query's fingerprint, silently under-counting N+1

[KARTIKrocks]

Problem

analyzer.Redact resolves the backslash-escape dialect ambiguity by scanning
under both readings and redacting the union (see PR #61). That is the right
fail-safe direction for a leak, but it has a consequence beyond readability:
an over-redacted query gets a different fingerprint.

Fingerprint(`SELECT * FROM t WHERE p = 'C:\' AND id = 5`)
// → "SELECT * FROM t WHERE p = ?"

Fingerprint(`SELECT * FROM t WHERE p = 'plain' AND id = 5`)
// → "SELECT * FROM t WHERE p = ? AND id = ?"

Fingerprint is the N+1 grouping key (middleware/n_plus_one.go) and the
dedup key (middleware/guard.go). So one logical query whose values only
sometimes contain a backslash — Windows paths, LIKE ... ESCAPE patterns,
regexes — groups under two fingerprints instead of one.

Impact

• N+1 detection can silently miss. A 20-iteration loop can land as 12 + 8
  and fall under a threshold of 20 on both, reporting nothing.
• De-duplication emits the same finding twice per window, once per group.
• Metric-label cardinality grows with a value-dependent property, which is
  exactly what Fingerprint is documented to avoid.

This is a false negative in a detector, which is the failure mode the repo
treats as worse than a hard error elsewhere (see the MySQL EXPLAIN
fail-closed behaviour in 0.1.1).

Current state

Documented, not fixed — website/docs/redaction.md ("When the dialect is
ambiguous, it over-redacts") and the ## [Unreleased] CHANGELOG entry both
call it out, with bind parameters as the workaround. AGENTS.md records it
as the accepted cost of the union scan.

Possible directions

• Narrow the union to the divergent tail, so the spans agree wherever the two
  readings agree and only the genuinely ambiguous region widens.
• Derive the fingerprint from the narrow reading while Result.Query keeps
  the union. Needs care: the fingerprint must stay PII-free, so this only
  works if the narrow reading provably leaks nothing into the folded form.
• Accept it and make it observable instead — e.g. mark a Result whose
  redaction was ambiguous so operators can spot skewed grouping.

Acceptance criteria

• Two executions of the same query shape share a fingerprint whether or not a
  value contains a backslash, or the limitation is deliberately kept with
  a documented, tested rationale.
• No regression in TestRedactNoLeakAcrossDialectAmbiguity — no literal byte
  may survive Redact or Fingerprint under either dialect reading.
• N+1 grouping behaviour covered by a test at the middleware level.

References: #61