[ramondalmau]

Problem (audit D4)

Operators can send their own fuel flow, aircraft mass, engine efficiency or true airspeed per
trajectory point, and may leave some values empty. PyNEATS resamples every trajectory to one
point per minute by drawing a straight line between neighbouring operator points.

An empty value broke the two lines touching it: every minute between the previous and the next
operator point got no value. Those minutes counted as zero fuel, with no warning.

Example (golden flight EAF2143, operator points about every 5 minutes, fuel flow empty at 19:12):

Time	If the value had been sent	Before this fix	With this fix
19:02	0.685 kg/s	0.685	0.685
19:03	0.664	missing	0.676
...	...	missing (15 min)	...
19:17	0.530	missing	0.537
19:18	0.491	0.491	0.491

Total fuel: 3,988 kg with the value; 3,505 kg before this fix (-12.1 %); 4,051 kg with this fix
(+1.6 %: a straight line cannot know the dip at 19:12). An empty mass value: about -15 % before.
With dense operator data (a point every ~25 s) one empty value usually changed nothing.

Also: a column empty at every point, in a DataFrame given directly to FlightRunner or the
parser, was used as all-empty, so the flight got no fuel at all.

Fix (decision: Ramon Dalmau)

An empty value is skipped, as if that point had not been sent: each minute is interpolated
between the nearest valid operator values; before the first / after the last valid value the
nearest one is used. A column empty at every point is treated as not provided (computed by
PyNEATS). Gaps are filled by a straight line, not by BADA (no mixing within a column).

Effect

• NM sample (4,959 flights): bit-identical. These flights carry no operator performance data.
• Golden test data: 57 flight records change, exactly those whose inputs contain empty operator
  values (near the end of the flight). 47 regain their last minute of fuel (fuel and CO2
  +0.0002 % to +0.96 %); 10 change only efficiency- or speed-dependent results. Every other
  record is byte-identical.

Checks

• 6 new tests (tests/test_interpolation_partial_nulls.py); 5 fail on the old code.
• Full test suite: 337 passed. Golden pipeline tests rerun after a final cleanup: 25 passed.
• Independent review: no code defects; its wording corrections are applied.
• Docs: input-format.md explains empty values; audit D4 closed with the numbers above.

Known limit: an empty mass value at the very start or end is held constant, so no fuel is
derived there (same total as before).