test: close the coverage gaps outside the transformer (98.94% -> 99.25%)

[tschm]

Follows #422. Together the two take the suite from 98.94% to 99.94%, with three lines left deliberately.

Fifteen lines across four modules. Each was picked because it is shipped behaviour that nothing held, not because it was a branch that happened to be missed.

MemoryBlobStore had no test at all

This is the one worth reading twice. MemoryBlobStore is exported from loman.serialization, its own docstring calls it "a worked example of the interface", and docs/paper/loman.tex:652 names it alongside DirBlobStore and ZipBlobStore as one of the three stores that ship.

grep MemoryBlobStore tests/ returned nothing. The suite has always brought its own store — RecordingStore in test_byos.py, written deliberately "the way a user would implement one" — so the store users are actually pointed at was never exercised, including the SerializationError its read_blob raises on a missing key.

Now covered: construction with and without a backing dict, the round-trip, the missing-key error, and a real end-to-end save/load routing a node's values through it.

The rest

• _require_parent — saving under a directory that does not exist names that directory. The 0.7.0 changelog promises this, and specifically promises the error does not name the sibling .tmp the containers build into. Both halves are now asserted, for the zip and dir containers.
• Named timezones — == on two aware datetimes compares instants, so a round-trip that quietly replaced Europe/London with a fixed +00:00 would still have passed the AWKWARD_VALUES table. The new table entry covers the encode and decode paths; three explicit tests assert the zone key survives, which is the part equality cannot see.
• put_blob with raw bytes — both the bytes and writer-callable forms are documented on put_blob; only the callable one was reached, because that is what the built-in transformers use.
• write_dill_old — applying the __serialize__ tag rule, and writing to an open file object. Worth knowing: the deprecated path implements that rule a second time, because it has deleted __getstate__ for the duration of the write. Only the other copy was covered, so the two could have drifted silently.

Result

	Before	After this PR	With #422 too
computeengine.py	99%	100%	100%
serialization/blobs.py	95%	100%	100%
serialization/values.py	98%	99%	99%
Suite total	98.94%	99.25%	99.94%
Uncovered lines	51	36	3

The three lines left, and why

ui/widget.py:947 (_select_key returning early with no view), ui/widget.py:1008 (the build-request early return) and serialization/values.py:485 (_reapply_freq returning unchanged when freq is None).

All three are reachable, so 100% is available. They are left because a test for each asserts little beyond "the guard is a guard", and because taking the gate to --cov-fail-under=100 has a cost worth deciding separately: this codebase carries 35 considered # pragma: no cover markers, and a 100% bar turns every future defensive branch into a choice between a tautological test and a pragma. 99.94% with three honest gaps reads truer than 100% leaning harder on suppressions. Happy to close them if you'd rather have the round number.

Verification

Gate	Result
make fmt	PASS — 21/21 hooks
make test	PASS — 1557 passed, coverage 99.25%