Non-booking emails can still yield an invented flight number

[Abrechen2]

What a user sees

Importing an email that contains no booking — a fare advertisement, an airline
status notification — can create a flight with a made-up flight number. The
entry looks ordinary in the logbook, so it is easy to accept and hard to spot
later.

Status

The common shape is fixed on fix/forgejo-round-2 (not yet merged): German
price copy such as "ab 380 EUR" was read as flight AB380, and "Nur 7 …" as
NUR7. A guard for this existed but never fired — the candidate reached it in the
source's original case, so "ab" never matched the listed "AB", and only the
first two characters were compared, which cannot see a three-letter word. Both
holes are closed and covered by tests.

What is still open

Two shapes remain, and neither can be closed by extending the word list,
because each collides with a genuine airline code:

1. An airline status email (no booking in it) yields DL3. DL is Delta.
2. A real booking confirmation yields MO24, apparently from a German
   weekday and day-of-month ("Mo, 24."). MO is likewise assigned.

Banning either prefix outright would drop real bookings, so this needs
context — the surrounding words, a label, a nearby route — rather than a
blocklist.

A second defect, found alongside

On that same genuine confirmation the real flight numbers are missed entirely;
the invented one is what remains. So this is not only a false positive but also
a miss, and the two probably want fixing together.

Why the evidence rule does not catch these

The rule added for the earlier report accepts a candidate that carries a flight
number or a full route. An advertisement manufactures exactly a flight
number, so it satisfies the rule. Anyone reading that earlier fix as closing
this whole class will be misled — it does not.

Notes

Found by running the parser over an offline corpus of real booking and
marketing emails. That corpus belongs to a third party and is not part of this
repository; the reproductions above are described rather than quoted, and no
sample is attached.

[Abrechen2]

Partly addressed on fix/forgejo-round-2 (not merged). Measuring the two cases
turned up a shared cause that also explains a second, opposite fault: the
same confirmation that produced MO24 was losing its real flights.

The context lookahead used ., which does not cross a newline. A confirmation
prints its flight number alone on a line, so nothing followed it on that line and
the real flight never matched. The line below — weekday, date, and an airport
name containing "Flughafen" — satisfied the lookahead instead. The rule meant
to demand context was preferring prose lines to flight lines. Selection made it
worse: the first hit in the document won, and the fare is printed above the
itinerary.

Now every match is collected and one whose prefix the airline catalogue knows is
preferred — as a preference, not a requirement, so an airline the catalogue has
not heard of can still produce a flight.

Measured over 337 archived mails. Recovered, each previously a phantom or
nothing: EK0050 (×8, across confirmations and their forwards), UA384,
UA252, UA102, LH2040, LH109, EW2755, DI7006, AF1423. Dropped:
DL3 — the case reported here — plus five marketing phantoms and the PRO1
series. Mails yielding no flight number rise 162 → 200.

Still open, which is why this issue stays open:

• Two marketing mails move from nothing to a phantom (IYR724, FB23), and
  several swap an obvious phantom for a more plausible one (MD2400). A code can
  be a real airline and a coincidence at once; the evidence rule cannot help,
  since a flight number is exactly what these carry.
• MO24's sibling: airberlin's code is AB, which the false-prefix list must
  ban because German price copy writes "ab 89 EUR". Its confirmations therefore
  yield nothing rather than their real flights. That trade predates the change —
  it is simply no longer hidden behind a phantom.