Preserve default behavior and expose unimem without std

Review-ready; production unchanged at 9cbd48070a136e24711896c930670fcff8a147fc. Serves Kadek phase 1 / native Image dependency. Fresh different-vendor PLAN review and substantive finding resolution precede production; fresh CODE approval precedes ordinary source acceptance. No version, tag, release or publication.

1. Freeze the exact parent, existing native tests, compile-only consumer and crossbeam feature/source inputs → verify the existing driver builds; the explicit no-default consumer fails E0152, while its ordinary std control builds. Keep initial newer-cache resolution separate from the final workspace-pinned consumer. Preserve literal consumer source and lock for the candidate. The parent already passes 53 tests + 13 doctests and 44 isolated native tests; the standalone no_std consumer and proposed durable gate both fail E0152.
2. Give unimem default = ["std"] and std = ["crossbeam-queue/std"]; request crossbeam-queue with default features disabled and alloc enabled at its workspace declaration, retaining version requirements/lock → verify default dependency features preserve the prior std-enabled queue/utils behavior, and the isolated no-default target closure contains no runtime std dependency. Cargo feature unification can re-enable std when another consumer requests it; document this.
3. Add #![no_std], use core for existing runtime namespaces, and import test-only std/prelude items explicitly → verify no changed FFI signatures, safety contract, allocation algorithm, public types or native layout; the frozen no_std consumer now compiles with the same source/lock. Its inert allocator remains compile-only. Add test-only extern crate std and explicit thread_local, vec, Vec imports; preserve existing test bodies/events. Integration tests/examples are separate std crates. Runtime unimem itself needs no std namespace import.
4. Install the frozen Nu gate and three fixture files → verify parent E0152, candidate pass, checksum-drift rejection before Cargo, and unchanged workspace/ template locks. Derive paths from the script; respect caller Cargo/toolchain/ CARGO_HOME. Compile in one temporary target, clean only that directory, never execute the inert allocator/driver. Compare locked queue/utils/mutants identities, unimem version and normal dependency names. Offline cache misses remain real prerequisite failures; no silent pin updates. Add the named gate to CLAUDE.md.
5. Run native owning tests/docs under default and no-default features, the existing native examples/benches, fmt and strict workspace Clippy/build/tests → retain actual commands, source epochs and any pre-existing broader failures. Bench execution is not a new numerical performance guarantee. Do not silently replace native calls with mocks to claim a no_std runtime result. Also run the isolated 44 native tests with identical bodies/lock and only the dependency path changed. Owning no-default tests can unify std through reverse dev-dependencies and are insufficient alone. Retain default, explicit std, isolated no-default and owning test feature trees. Compare actual std/alloc cfgs; the unused queue default marker can disappear. Source inspection establishes its lack of cfg consumers.
6. Check public API from separate parent/candidate rustdoc JSON and actual last tag availability; record additive feature policy and reverse consumers → preserve real gate gaps. Fresh CODE review sees the complete small delta, exact controls and compact audit. Commit/push an ordinary PR under standing owner authorization; no Image phase completion or delivery claim follows. Select the explicit unimem package to avoid preexisting audit manifest ambiguity. Preserve actual disjoint compilation roots/input identities. The exact tag gate already fails at the historical NeBu path and stays unmet; a repaired tag harness is no substitute.

Files: root Cargo.toml (queue feature defaults only), unimem/Cargo.toml, unimem/src/{lib,block,ffi,tape,grid,layout}.rs, block/creation.rs and the minimum test-only imports in block/tests.rs. Existing tests keep their expectations. One concise README usage note explains Apple-only no-default operation and the allocator/feature-unification boundary; canonical specifications remain unchanged. Additional files: unimem/CLAUDE.md, unimem/scripts/check-no-std.nu, and unimem/fixtures/no-std/{consumer.rs,Cargo.toml.txt,Cargo.lock.txt}. The templates are never repository Cargo manifests. Frozen proposed bodies are in /tmp/unimem-core-gate-draft/unimem/; no production copy is installed before review. Research, reviewed plan and observations stay in task/audit directories. Keep replay tools in Nu/Rust; temporary historical capture helpers are evidence.

Five likely failures and their checks: transitive std remains (own panic-handler consumer); default queue backoff changes (feature graph equality); test prelude breakage (both feature-mode native suites); namespace replacement changes public traits or FFI (full diff/API comparison); no_std gets confused with platform or native ownership closure (explicit remaining requirements in research/report).

The core-only feature mode may spin under queue contention where std yields; that is the dependency's explicit no_std policy, not default-mode equivalence. Grid remains an allocating legacy utility, with no new memory/work bound. Block still exposes raw capabilities; no alternate storage type or synchronization layer is introduced. If a real gate requires a wider semantic change, stop this source unit at the measured boundary and bring that change into its own reviewed plan.

Required candidate commands (locked/offline where Cargo supports them): owning build, test default/no-default, standalone durable check, isolated native tests, bench, release pipeline; workspace fmt --all --check, strict Clippy, release build, tests; explicit parent API comparison. Use separate parent/candidate compiler targets, jobs 2/incremental off, declared compiler/warning flags. Preserve original baseline records 01–17 including preparation errors. Repeat an affected gate only for a source/failure/input change. No performance threshold or benchmark improvement claim is introduced by this namespace/feature boundary.

Computed floor: additive feature policy at least minor/ring 1. Registry Honeycrisp belongs to Nox; concrete reverse consumers include Cyb and the pinned manifest inventory in research. Rung product/dependent/specification/Kelvin gates remain unverified, tag comparison blocked, workspace fmt/Clippy/tests red. Keep decision and any required wip: first-error commit wording under standing ordinary-source authorization. No dependency versions, Cargo.lock, sibling pins, FFI signatures, allocation/access semantics or release actions change. Final CODE review receives these limitations, not an inferred all-green or full-native-Image claim.