Tests: cover the security-critical paths of the crypto and platform layers.

[bernardladenthin]

What

Adds 14 test blocks to the existing self-test suites. No production code is touched and no
existing test is modified — 1100 lines, additions only, across 6 files:

File	Blocks added
src/Volume/EncryptionTest.cpp / .h	8 — API rejection paths, hash wrappers, KDF selection, volume-header rejection, password comparison and cache, keyfile mixing, VolumeInfo round trip, volume layouts
src/Platform/PlatformTest.cpp / .h	4 — Buffer/SecureBuffer/Memory, StringConverter, file operations, exception transport
src/Main/UserInterface.cpp / .h	2 — random number generator, MountOptions round trip

The tests target paths the existing known-answer tests never reach, because those only ever
take the success path: rejection of bad parameters, uninitialised state, out-of-range access,
and the serialisation round trips that cross the IPC boundary to the privileged core service.

Rebased

Rebased on master @ fe33e447 after #1844 was merged, without conflicts. Since this PR was
opened, CoreTest::VolumePasswordTest arrived upstream and covers the password size limit and
the serialised round trip, so those two parts of TestPasswordHandling were removed instead of
testing the same thing twice. What remains there is the comparison of two different passwords
and the password cache. All coverage figures and the password rows of the detection proof
below were measured again on the rebased branch.

Coverage

Measured A/B in a single binary — same build, the new test calls commented out for the
baseline — using gcov. Base is master @ fe33e447.

File	before	after
Volume/VolumeLayout.cpp	0.00 %	95.14 %
Volume/VolumePasswordCache.cpp	0.00 %	94.44 %
Core/RandomNumberGenerator.cpp	0.00 %	84.51 %
Volume/Keyfile.cpp	0.00 %	57.73 %
Volume/Hash.cpp	13.54 %	96.88 %
Volume/VolumeInfo.cpp	3.19 %	73.40 %
Platform/FileCommon.cpp	23.26 %	86.05 %
Core/MountOptions.cpp	27.35 %	85.47 %
Platform/Memory.cpp	79.41 %	91.18 %
Volume/VolumePassword.cpp	92.16 %	98.04 %
Volume/Pkcs5Kdf.cpp	50.48 %	70.48 %
Volume/Cipher.cpp	73.24 %	83.57 %
Platform/Buffer.cpp	77.78 %	83.84 %
Volume/EncryptionAlgorithm.cpp	78.28 %	93.43 %
Volume/VolumeHeader.cpp	81.90 %	83.49 %
Volume/EncryptionModeXTS.cpp	91.39 %	93.38 %
Platform/StringConverter.cpp	40.41 %	57.96 %

17 files improved, four of them starting from zero. The tests that arrived upstream in the
meantime raised several baselines (VolumePassword.cpp, Memory.cpp, Hash.cpp).

Coverage was produced without modifying any build file, using the existing hooks in
src/Makefile:

make NOGUI=1 NOTEST=1 \
     TC_EXTRA_CFLAGS="--coverage -O0" \
     TC_EXTRA_CXXFLAGS="--coverage -O0" \
     TC_EXTRA_LFLAGS="--coverage"

A note for anyone repeating this: use gcov directly. Both lcov and gcovr mis-attribute
results here, because each subproject is built through make -C and the working directory
recorded in the .gcno files cannot be resolved afterwards. lcov reported one file at 8.7 %
where the true figure is 91.4 %, and produced merged totals lower than the individual runs.

Detection proof

Coverage alone does not show that a test would catch anything, so every block was verified by
injecting a defect and confirming that a test reports it. 27 injected defects when the PR was
opened; the password defects were injected again after the rebase. Among them:

Injected defect	Reported by
Buffer::Erase no longer wipes	BufferTest
Buffer::GetRange skips its bounds check	BufferTest
Memory::Compare inverts its ordering	BufferTest
keyfile contributes nothing to the password pool	TestKeyfileApplication
base password dropped from the keyfile pool	TestKeyfileApplication
empty keyfile silently accepted	TestKeyfileApplication
MountOptions loses the protection mode in transit	TestMountOptionsSerialization
password truncated in transit	CoreTest::VolumePasswordTest (upstream; the duplicate block was removed)
password comparison ignores the last byte	TestPasswordHandling
password cache grows without bound	TestPasswordHandling
password cache stops deduplicating	TestPasswordHandling
RNG returns constant data	TestRandomNumberGenerator
RNG usable before Start()	TestRandomNumberGenerator
oversized RNG request not rejected	TestRandomNumberGenerator
File::Copy truncates the copy	FileTest
ReadCompleteBuffer accepts short reads	FileTest
exception subject dropped on the IPC path	ExceptionTransportTest
cipher accepts block operations before a key is set	TestEdgeCases
SetKey accepts a wrong-length key	TestEdgeCases
hash accepts empty input	TestHashClasses
KDF lookup matches any name	TestKdfSelection
volume layout reports a data area past the end of the host	TestVolumeLayouts

One side effect worth mentioning: the random number generator carries its own pool-mixing
known-answer self-test inside Start(). That self-test previously ran only when the graphical
or the text interface started the generator — --test never started it, so it never ran under
CI. It does now.

Verification environment

Linux — when the PR was opened: Docker ubuntu:24.04, GCC 13, distribution wxWidgets 3.2,
make NOGUI=1. After the rebase: Docker debian:trixie (amd64, GCC 14) for the build (0 warnings),
three --test runs and the coverage and defect runs above; and Alpine 3.22 i386 (musl) with this
branch merged with the pending i386/sparc64 build fixes: 0 warnings, three --test runs. The
repository is mounted read-only and copied into the container before building, so no build
artefacts touch the working tree. The self-tests were run twice to confirm repeatability, and
the temporary files created by the keyfile and file tests are removed even when an assertion
fires.

Windows — when the PR was opened, Visual Studio 2022 17.14.37, MSVC 14.44.35228, Windows SDK 10.0.26100.0,
WDK 10.0.26100. Each affected translation unit was compiled in user mode at /W4 and in
kernel mode at /W4 /WX, matching the TreatWarningAsError setting in Driver.vcxproj.

None of the six changed files appears in any .vcxproj, so the Windows build cannot be
affected by this change. The compile checks were run regardless.

Limitations

• Volume/VolumeHeader.cpp moves by only 1.6 pp. Its remaining uncovered part is the
  decrypt-attempt loop, which needs a real encrypted volume — out of scope for a unit test.
• Volume/Keyfile.cpp stops at 58 %. The remaining branches are the security-token and
  directory-enumeration paths, which need a token or a populated directory.
• The repository has no Windows CI job, so the Windows half of these tests runs only when
  Format, Mount or ExpandVolume actually start, not automatically.

AI assistance

These tests were written with the assistance of Anthropic's Claude Opus. Every claim in this
description — the coverage figures, the injected-defect results, the build and warning output
— was produced by running the builds and measurements described above rather than estimated.
The changes are additive and confined to test code; review of the assertions themselves is of
course still welcome.

Assisted by Claude Opus 5.5

[bernardladenthin]

Please merge #1844 before this PR. Both touch src/Volume/EncryptionTest.{cpp,h}, and while
the changes are independent (different functions, no overlapping assertions), applying them in
the other order requires a three-way merge — a plain git apply of #1844 will fail against
this branch's header. Merging #1844 first avoids that entirely; I will rebase this branch on
the updated master afterwards.