core: one context struct for the (conn, root, idgen, cfg, embedder) bundle threaded through 15 functions in three orders

[samkeen]

Found during the pure-refactor pass on claude/rust-codebase-refactor-7ot3v1; deferred there because it's a large, test-touching diff that deserves its own PR — this issue is the map.

The smell

Fifteen function definitions in b2-core thread the same 4–5 values, in three incompatible parameter orders:

order	functions
conn, vault_root, path, idgen, cfg[, embedder]	ingest::ingest_file, ingest::project_file
conn, vault_root, idgen, cfg[, embedder]	ingest::project_vault, ingest::ingest_vault_with_progress, rm::reproject_dangled
conn, idgen, cfg[, embedder], vault_root	mv::move_note / move_resource / move_dir, add::add_note / create_note, rm::delete_note / delete_resource / delete_dir

Roughly 59 call sites carry the bundle (13 in vault.rs, ~30 in crates/b2-core/tests/), and four #[allow(clippy::too_many_arguments)] (add.rs ×2, mv.rs ×2) exist purely to silence it. rm::reproject_dangled even uses a different order from the three public fns in its own file.

The shape of the fix

Two context structs, split along the model-free vs. embedding posture the module docs already document (rm/create_note/write are model-free; mv/add/link embed):

struct ProjectionCtx<'a> {
    conn: &'a Connection,
    root: &'a Path,
    idgen: &'a dyn IdGen,
    cfg: &'a ChunkConfig,
}
struct EmbedCtx<'a> {
    proj: ProjectionCtx<'a>,
    embedder: &'a dyn Embedder,
}

Vault builds them via fn ctx(&self) / fn embed_ctx(&self), so a façade call collapses to e.g. mv::move_note(self.embed_ctx(), &b2id, &old_rel, to). The type system then encodes the posture — a model-free op cannot receive an embedder — which today is doc-enforced only. Deletes all four clippy allows and the ordering hazard.

Short-lived, Copy-able borrow-only view structs passed into calls and never stored are the sanctioned exception to the "prefer owned fields" rule (CLAUDE.md, Rust data modeling — the NoteRow precedent).

Riders worth taking in the same PR (same files, same review)

From the same survey of mv.rs:

• the move "preflight dance" (normalize → same-path check → occupied-target check → create_dir_all + rename) is written three times (move_note/move_resource/move_dir);
• the "rewrite inbound files" loop is written three times, with move_dir inventing an let empty = BTreeMap::new() sentinel;
• the wikilink .md-convention replacement and the fragment-preserving resource replacement are each computed in two places — as pure functions they'd finally get direct unit tests (the #fragment logic currently has none).