v0.7.0

strided-rs v0.7.0

Breaking: placement is an explicit Exec token

Every entry that can fan out takes a strided_exec::Exec as its last parameter, by value.
The token carries the pool and a budget of slots; the callee splits it (fork, for_ranges,
limit) and never starts a task by any other route. No thread-local placement state exists
any more, so work that runs on a waiting thread cannot inherit another operation's pool or
budget. All workspace crates are released at 0.7.0.

• strided-exec (new crate). Pool and PoolBuilder (the only place a thread pool is
  started), Exec (serial, new(&pool, budget), budget, is_serial, reborrow, limit,
  fork, for_ranges, broadcast, external), and the Error and Refused types. A token is
  neither Clone nor Copy; slots are conserved across nesting; a serial token, a budget of
  one, or a single task runs inline and never starts the process-global Rayon pool. Without the
  parallel feature the same types exist and every entry runs inline. No Rayon type appears in
  the public API (scripts/check-no-public-rayon.sh).
• Exec::external. Enters the pool once and hands the budget to an outside library (a BLAS,
  a linear-algebra backend) as a plain width.
• One function per operation. The *_par permutation copies are merged into copy_into,
  copy_into_uninit and copy_into_col_major. strided-perm uses the budget it is given and
  never chooses a width, and one MINTHREADLENGTH decides when a loop is worth splitting.
• strided-basic, strided-fused, strided-kernel, strided-nn take Exec;
  strided-basic re-exports Exec, Pool and PoolBuilder, and strided-kernel re-exports
  them from strided-basic. strided-fused and strided-nn use the token through
  strided-basic.
• Rayon is a dependency of strided-exec only. strided-perm and strided-basic no longer
  depend on it. cargo deny check bans rejects a direct rayon dependency (normal, dev or
  build) anywhere else in the workspace, and bans faer, t4a-faer and spindle.
• strided-nn 0.7.0 (previously 0.1.1): the fused axis-0 softmax takes caller-provided
  scratch sized by executable tasks (scratch_len) and an Exec.

Migration

0.6	0.7
ExecContext::serial()	Exec::serial()
ExecContext::max_threads(n)	Exec::new(&pool, n)? with a strided_exec::Pool
ExecContext::run_on(pool, ..)	build a Pool once, mint Exec::new(&pool, n)? per call
plan.execute(&ctx, ..) (context first)	plan.execute(.., exec) (token last)
with_execution_policy, ExecutionPolicy	removed; pass an Exec
with_execution_pool	removed; pass an Exec
rayon_threads()	exec.budget()
copy_into_par, copy_into_uninit_par, copy_into_col_major_par	copy_into, copy_into_uninit, copy_into_col_major with an Exec
ambient (global Rayon pool) placement	none; use Exec::serial() or a pool you own
StridedError::InvalidThreadBudget	strided_exec::Error::ZeroBudget from Exec::new
MaybeSend, MaybeSync bounds	Send + Sync
a direct rayon dependency for placement	strided-exec (feature parallel, default)

A phase sequence reuses one token with exec.reborrow(), because every parallel entry consumes
its token. Plans never store an Exec; replay takes it per call.

Design record: docs/design/cpu-exec-placement.md.

Full comparison: v0.5.0...v0.7.0