chore(worker): extract the Tokio data-plane path into tokio_conn

[beinan]

Refs #306. First of the splits proposed there — the one with a real seam.

Why this file, this cut

crates/talon-worker/src/main.rs was 1509 lines covering CLI parsing, backend construction, io_uring capability probing, ring supervision, coordinator registration, heartbeat, signal handling, and the entire Tokio data-plane connection handler.

The connection handler is the piece that clearly does not belong there, and the codebase already says so. uring_conn.rs is a library module whose own docs open with:

This is the completion-based counterpart to the Tokio handle_conn in main.rs.

So two implementations of one wire protocol lived at different layers — one in the library, one inline in a binary. They must stay behaviourally identical (same frames, same limits, same error semantics), and that is much easier to review and to keep honest when they sit side by side as siblings.

Change

Moved to a new crates/talon-worker/src/tokio_conn.rs:

• handle_conn, handle_control_frame, handle_put, handle_delete, sendfile_payload, read_control

Only handle_conn and read_control are pub — the other four stay private so the library's public API does not widen.

main.rs: 1509 → 1059 lines.

The module doc explains why this path pays for sendfile (Tokio's TcpStream is non-blocking, so a blocking sendfile would spuriously EAGAIN, forcing an into_std/from_std round-trip on every transfer) and points at uring_conn for the contrast, mirroring the existing cross-reference in the other direction.

This is a pure move

The function bodies are byte-identical. Verified mechanically:

$ diff <(git show HEAD:.../main.rs | sed -n '617,1066p') <(tail -n +38 .../tokio_conn.rs)

differs only in the doc-comment line offset and the trailing newline. The remaining diff is imports: removed from main.rs what moved out, added to tokio_conn.rs what it needs, and pushed four now-test-only imports into the mod tests block.

No behaviour change, so the existing suite is the regression test.

A gap this exposed

Worth flagging, because it is the kind of thing #306 argued splitting would surface. With the path in its own file, coverage attributes it separately:

uring_conn.rs   78.68%   <- the default since #299
tokio_conn.rs   36.41%   <- the fallback
main.rs         55.04%

The two paths are supposed to be behaviourally identical, so 36% on the fallback is a real gap — it was previously averaged into main.rs and invisible. Not addressed in this PR; this change only makes it measurable. Worth its own issue.

Verification

$ cargo build --workspace --all-features    # no warnings
$ cargo clippy --workspace --all-targets --all-features -- -D warnings   # clean
$ cargo test --workspace --all-features     # 36 suites, all ok

Not included

The other splits in #306 (runtime.rs, coordinator/observability.rs) are separate mechanical changes and belong in separate PRs. I have deliberately left talon-core/src/config.rs alone — as noted in the issue, it is the source of truth for the generated configuration reference and splitting it risks the config-docs drift gate for little gain.

---

Update: PR title changed refactor: → chore: (the repo's conventional-commits gate allows feat|fix|enhance|test|doc|bench|build|ci|chore, not refactor), and fixed a self-inflicted docs failure: the module docs linked to sendfile_payload, which this PR makes private, tripping -D rustdoc::private_intra_doc_links.