v0.9.0 — prax generate now emits a runtime-ready client

Added

• prax generate now emits a runtime-ready client. Generated
  model modules carry the trait impls the runtime needs to actually
  run queries, matching the surface produced by #[derive(Model)]:
  • impl prax_query::row::FromRow decodes scalar columns via
    FromColumn and default-initializes relation fields, so
    find_many and friends round-trip rows back into the generated
    structs at runtime.
  • impl prax_query::traits::ModelWithPk exposes pk_value() and
    get_column_value() for nested writes, upsert, and composite
    primary keys.
  • The per-model operations struct is named Client<E> (was
    {Name}Operations<E>), so prax_orm::client!(User, Post, ...)
    can resolve <snake_name>::Client<E> by path the same way it
    does for the derive path. The full CRUD surface — find_many,
    find_unique, find_first, create, create_many, update,
    update_many, upsert, delete, delete_many, count — is
    emitted on Client<E>.
  • Non-list relation fields are emitted as Option<T> (or
    Option<Box<T>> when boxing is needed to break a cycle)
    regardless of the schema modifier, so the FromRow default-init
    has a None to write into. The relation executor populates
    Some(T) on the .include path.

Fixed

• prax-migrate — CREATE TABLE emission respects FK dependencies.
  Before, SchemaDiffer populated create_models by iterating a
  HashMap, leaving the resulting CREATE TABLE order
  non-deterministic. A schema where tracks and playlists
  reference sync_sources could emit sync_sources last; SQLite
  tolerated it because FK targets are resolved at row-write time,
  but strict engines (Postgres, MySQL with FK enforcement, MSSQL)
  and any deferred-constraint bootstrap would fail to apply the
  migration. SchemaDiff::ordered_create_models now does Kahn's
  algorithm topo sort over the FK graph: referenced tables emit
  before their dependents, self-references and FKs that point at
  out-of-batch tables don't constrain ordering, and cycles fall
  back to original order. All five SQL generators (Postgres,
  MySQL, SQLite, MSSQL, DuckDB) route through it and emit drops
  in the reverse direction so rollbacks drop dependents before
  parents.

Changed

• Workspace clippy gate is back online. The husky pre-commit
  hook had silently been bypassed in environments that override
  core.hookspath with no project-local pre-commit; develop had
  accumulated 200+ clippy errors under -D warnings. Cleared every
  diagnostic so
  cargo clippy --workspace --all-targets --all-features -- -D warnings
  passes again. API-shape lints (result_large_err,
  new_ret_no_self, should_implement_trait, pedantic noise in
  prax-scylladb) are suppressed at crate level with rationale;
  bug-shaped lints (manual_checked_ops, manual_strip,
  manual_clamp, manual_sort_by_key, &PathBuf → &Path,
  mixed_attributes_style, unnecessary_unwrap) are fixed in
  place.
• prax-sqlite — vector tests skip cleanly when the loader is
  unconfigured. The two unit tests in vector/register.rs and
  the three integration tests in tests/vector_integration.rs no
  longer fail under cargo test -- --include-ignored (used by CI)
  in environments that haven't provisioned the sqlite-vector-rs
  cdylib. They detect the missing library at runtime via
  SQLITE_VECTOR_RS_LIB and bail out with a "skipping" message
  instead of panicking on the loader error.