fix(sphragis)!: make deterministic encapsulation private

[forkwright]

What

EncapsulationKey::encapsulate_deterministic accepted caller-supplied KEM
randomness and was pub (gated only by #[doc(hidden)], which hides
rustdoc output, not visibility). Any consumer of sphragis could call it
directly, despite its own doc comment warning "never call with non-uniform
or reused randomness." Reused or predictable input deterministically
collapses the ephemeral X25519 secret, the ML-KEM coins, the ciphertext,
and the shared secret — the exact property EncapsulationKey::encapsulate
exists to guarantee.

• Dropped pub and #[doc(hidden)]: encapsulate_deterministic is now a
  private inherent method. It stays plain-private rather than gaining a
  cfg(test) gate, because encapsulate() calls straight through to it
  with fresh OsRng bytes in every build — the method must compile
  unconditionally, and privacy alone already makes it unreachable from any
  downstream crate (a tests/ integration file compiles as a separate
  crate, same as a real consumer).
• Relocated the X-Wing draft known-answer test beside the implementation,
  into src/hybrid.rs's own #[cfg(test)] mod tests — it's the only test
  build with visibility into the now-private method.
• Added a compile_fail doctest on EncapsulationKey (the fleet-documented
  pattern for API-surface guarantees, basanos/standards/RUST.md § Compile-Fail
  tests) that calls encapsulate_deterministic from outside the crate and
  asserts it does not compile — an executable assertion, not a comment.

Closes #17

Why private, not hazmat feature

No real external conformance consumer exists: the only caller was this
crate's own KAT (tests/known_answer_vectors.rs), and akroasis — the sole
declared consumer of sphragis — doesn't reference encapsulate_deterministic,
or in fact anything from this crate: it names sphragis in
[workspace.dependencies] but no member crate depends on it and no .rs
file in the tree references it at all (verified by tree-wide grep). A
hazmat feature would still ship the
method in the published crate, reachable by any consumer that flips one
flag; plain privacy makes it unreachable by construction. If a real external
KAT/conformance use case appears later, the issue's own fallback path
applies: an explicitly unstable hazmat feature/module with a distinct
type and a misuse-resistant contract — not a flag on the production key
type.

Verification (fails before, passes after)

Before this change, tests/known_answer_vectors.rs called
ek.encapsulate_deterministic(&eseed) directly as an external crate
(everything under tests/ compiles separately from src/, exactly like a
downstream consumer) — and it compiled and passed, which is the concrete
proof of the reachability this issue reports. That call is now gone from
tests/; left in place, it fails to compile with "no method named
encapsulate_deterministic found" once the method drops pub. The new
compile_fail doctest on EncapsulationKey pins this down permanently: it
must fail to compile, and does — a regression that re-exposes the method
(e.g. someone reverting to pub) fails cargo test --doc --features preview-pq in CI (gate-attestation.yml's doctest_cmd).

cargo fmt --all -- --check
cargo check --all-targets && cargo check --all-targets --features preview-pq
cargo clippy --all-targets -- -D warnings && cargo clippy --all-targets --features preview-pq -- -D warnings
cargo test && cargo test --features preview-pq

Breaking change (pre-1.0, semver-minor)

EncapsulationKey::encapsulate_deterministic is removed from the public
API surface. sphragis is 0.1.2, pre-1.0. akroasis pins its
[workspace.dependencies.sphragis] to tag = "v0.1.1" — already two
releases behind current main — so it will not take this break on any
ordinary resolve; picking it up requires someone to deliberately move the
pin. Named and verified blast radius today: zero. akroasis declares the
dependency but no crate in its workspace depends on it and no .rs file
references sphragis at all, let alone encapsulate_deterministic.

Standard traceability

No new algorithmic choice: X-Wing (draft-connolly-cfrg-xwing-kem) and its
combiner are unchanged; the deterministic-encapsulation code path itself is
untouched, only its reachability. The known-answer vector moved, not
edited — same seed, eseed, and expected shared secret as the RustCrypto
x-wing test-vectors.json entry [0] this crate has always gated on.

[forkwright]

Rebased onto current main (now carrying #24 bounded-CBOR, #27 KAT provenance lock, #32 public-API narrowing). src/hybrid.rs auto-merged textually (this branch and #32 touched non-overlapping regions); tests/known_answer_vectors.rs conflicted and was resolved by hand.

Does #32 subsume this PR? No. #32 narrowed the public API (HybridKem, SharedSecret, EncapsulationKey::encapsulate/DecapsulationKey::decapsulate, derive_wrap_key all moved behind hazmat) but left EncapsulationKey::encapsulate_deterministic exactly as this PR found it: #[doc(hidden)] pub fn, unconditionally reachable regardless of hazmat (EncapsulationKey itself is always-public, and #[doc(hidden)] hides rustdoc output, not visibility). This PR's privatization remains fully necessary.

#27's guarantee, carried through the relocation. #27 upgraded the X-Wing KAT from a shared-secret-only check against hardcoded hex literals to a full sk/pk/ct/ss check reading from the vendored, hash-locked vector file (crypto-provenance.toml: xwing-kat-0). Relocating the test into src/hybrid.rs's own #[cfg(test)] mod tests (required once encapsulate_deterministic is private) used #27's enhanced version, not this branch's original hardcoded-hex one — otherwise the relocation would have silently regressed #27's own guarantee. crypto-provenance.toml's executed_by field and review_policy.trigger prose updated to point at the new location instead of going stale.

Also completed while in this code: the SharedSecret type alias was left unconditionally pub by #32, with a doc comment explaining why — encapsulate_deterministic returned it as a pub (doc-hidden) method, so narrowing the alias would have leaked it through that signature (private_interfaces). That blocker is exactly what this PR removes. Gated SharedSecret itself behind hazmat too, matching HybridKem's existing pub(crate)/pub split, and corrected the now-stale doc comment.

CI green on the current head.