Close the last findings of the sixth review - #72

The completion handler stopped offering object properties the moment a character
was typed after the dot. ExtractObjectPath required the line to end with one, and
the editor re-requests completion on every keystroke — so the list appeared after
"Server." and vanished at "Server.P", exactly when it becomes useful. The
identifier being typed is now stripped before looking for the dot; filtering it
was already handled separately. This is the behaviour the previous round could not
characterise and removed a test for rather than guess at. The answer was that it
did not work.

Four invariants that mutation testing showed nothing protected now have tests:
short-circuit evaluation of and/or, which is written and commented but was only
observable when the right side would fail; duplicate let detection, a deliberate
validation with a user-facing message and no test anywhere; LF normalisation of
generated files, which would churn every diff on every platform; and
case-insensitive comparison of flattened array entries, which would otherwise
report a field as leaking when the override does redefine it in different casing.

Three documentation claims corrected against what the compiler does: output goes
to the current directory, not next to the source; commas are optional only across
newlines, since two items on one line still need one; and the specification's
"exactly one settings block" has been wrong since multiple blocks began to
deep-merge, which is what lets an included file contribute one.

409 -> 415 tests.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>