Review follow-up to #3777.

The fallback deliberately does not top up per entry: when items yields a
File for some entries and null for others, the null ones are dropped
rather than recovered from files. That is correct, because files mirrors
the same payloads, so a per-entry fallback would re-read a payload items
already yielded and attach it twice, which is the bug #3777 fixed. Losing
an unreadable entry is the right trade against re-introducing that.

Only a comment said so. The two existing fallback tests cover the
all-or-nothing extremes, so a contributor "fixing" the partial-null case
would find every test still green. This pins the decision.

No production change.

GitOrigin-RevId: df4bbb3ce4efb127c2b8195f507953fffdd35439