docs(email-security): body similarity groups by normalization, not by tolerance

[maximelb]

Why

The campaigns page's Body similarity section made two claims from single measurements and
they read as guarantees:

• "86 apart before normalization and 0 apart after it" — real, but one synthetic pair;
• a warning that a single unremovable per-recipient word "has been measured at 100" — also
  real, and also one pair.

Re-measured across a whole corpus rather than on one pair, the cost of one per-copy word the
normalization cannot identify has a median of 20–30 and a p95 over 100 (worst 201). So the
old text overstates the failure in the typical case and understates the variance, and — the
part that matters operationally — it leaves "raise the threshold" looking like a reasonable
response. It is not: moving 30 to the ceiling of 35 takes that case from roughly half to
roughly two thirds while spending most of the margin against unrelated mail.

Meanwhile the product side of this shipped five more normalizers (all recipients' names, wider
greetings, signature blocks, quoted-reply tails, short mixed letter-and-digit references, and
whole-number collapse), so the list of what the normalization removes was out of date and the
benign-corpus margin moved 39 → 40.

What changed

• The normalization list now describes what is actually removed.
• The measured example is now a kit templated over eight recipients in three shapes
  (individually addressed, shared mailbox, thread-hijack) — 37–219 apart raw, 0 apart
  normalized — rather than one pair.
• The warning admonition is rewritten so the threshold is not read as a tolerance dial, and
  says plainly that raising it is not the lever it looks like.
• 39 → 40 for the closest unrelated pair, in both places it appears.

No behaviour change here; this is the public wording catching up with the measurement.

Hold

Do not merge yet — this describes normalizers that are not in production. Merge with the
prod release, same as #388 / #389.

[maximelb]

/lc-review

[maximelb]

HOLD — do not merge yet. Approved and green, but this describes normalizers that are not in production: it depends on go-mailsec#134 and its legion_mailsec re-pin reaching the prod datacenters. Merge with the prod release, alongside #388 and #389.

[maximelb]

/lc-review

[maximelb]

/lc-review