Sanitization 2.0.2

sanitization 2.0.2

This patch release improves Miri compatibility and strengthens mapped-memory
teardown without changing the public API.

Miri lifecycle coverage

• Native locked containers use a private aligned-allocation model only under
  cfg(all(miri, test)), in the core crate's own unit-test build, instead of
  executing unsupported syscall inline assembly.
• The model covers fixed, dynamic, and UTF-8 locked storage, pool allocation
  and reuse, random-canary ownership, integrity failure, quarantine, rollback,
  growth, replacement, and drop.
• Every simulated mapping is checked to be entirely zero before deallocation.

Miri does not execute or validate real mmap, mlock, dump/fork policy,
CSPRNG, page-protection, or guard-page operations. Simulated successful report
states are test state only; native target evidence remains required. A normal
build with --cfg miri does not select the simulator, and a release build that
also forges cfg(test) is rejected. Downstream Miri uses portable comparison
code, while tests that execute native mapped constructors remain unsupported
and should target-gate those paths.

The same condition now protects every production comparison, AArch64
page-size, cache-flush, register-scrub, guard-page, and interop cfg boundary. A
release gate statically rejects Miri protection behavior switches without
test, compiles a normal release library with a manually supplied --cfg miri,
and rejects the forged release simulator combination. The release helper also
refuses ambient RUSTFLAGS and CARGO_ENCODED_RUSTFLAGS; evidence records both
channels. These checks treat the compiler invocation as trusted.

The full Miri workflow runs all-feature mapped lifecycle coverage as core
library unit tests and runs derive and companion integrations with portable
comparison features. This preserves their interpreter coverage without making
native production dependencies select a simulator.

Teardown hardening

LockedSecretBytes, LockedSecretVec, and SecretPool now clear the complete
mapping immediately before unlock and unmap. This includes mapping padding and
integrity metadata in addition to the secret payload.

SealedSecretBytes no longer unmaps pages when cleanup cannot confirm that
every page is erased. Cleanup now makes each page writable, erases it, and
reseals it immediately, continuing across other pages after a failure. It
retains the poisoned mapping, any uncertain page, and any established lock for
checked retry; Drop deliberately leaves that mapping to process teardown
rather than returning uncertain physical pages to the operating system.
Native fault-injection tests cover both ordering directions: a first-page
failure does not prevent later pages from being erased, and a cleanup reseal
failure retains already-erased storage until checked retry succeeds.

Mapped native and subtle equality traits now fail closed with a false choice
on integrity failure instead of selecting an implicit panic policy. Checked
try_constant_time_eq remains the API for distinguishing canary corruption
from ordinary inequality. Miri now compiles and exercises the locked dynamic
and UTF-8 zeroize/subtle interop implementations.

All five workspace crates are released together at 2.0.2, with the derive
crate exact-pinned to the matching runtime version.

The fuzz-only NCSA exception for libfuzzer-sys 0.4.13 is now held in a
fuzz-specific cargo-deny configuration, avoiding unmatched exceptions in the
runtime and tooling dependency graphs.