refactor(ops): decouple task-per-tx drivers from SubOperationTracker (#1454 phase 3c)

[iduartgomez]

Problem

After Phases 3a (#3843) and 3b (#3884) migrated client-initiated PUT and GET to task-per-tx drivers, both drivers still registered subscribe children with SubOperationTracker via expect_and_register_sub_operation. This was only needed so op_manager.is_sub_operation(tx) (a DashMap lookup) returned true at 4 guard sites that prevent spurious client notifications. The task-per-tx drivers never used the tracker's parking-lot mechanism — they await or fire-and-forget children directly.

Additionally, async non-blocking subscribe children (track_parent=false) were never registered in the tracker, so op_manager.is_sub_operation() returned false for them — they'd incorrectly try to notify a nonexistent client waiter at LocalSubscribeComplete.

Solution

Switch all 4 is_sub_operation guard sites from op_manager.is_sub_operation(tx) (DashMap lookup on SubOperationTracker.parent_of) to tx.is_sub_operation() (structural check: Transaction.parent.is_some(), set at creation by new_child_of). This is a superset — correct for both legacy and task-per-tx children, and fixes the latent bug for async non-blocking children.

Then remove expect_and_register_sub_operation calls from PUT and GET task-per-tx drivers, making them fully independent of SubOperationTracker.

SubOperationTracker itself stays alive — legacy PUT/GET relay paths still use the parking-lot mechanism. Full deletion deferred until relay migration (#3883) completes.

Guard sites changed

File	Context
p2p_protoc.rs:2061	LocalSubscribeComplete silent-absorb
node.rs:552	Subscribe completion — skip client notify for children
node.rs:646	Error path — skip client error for children
subscribe.rs:932	handle_abort — sub-op vs standalone abort routing

Testing

• cargo clippy -- -D warnings — clean
• cargo test -p freenet — 2380 tests pass (4 pre-existing macOS connectivity failures unrelated)
• Source-scrape test updated to reflect removed tracker registration

Fixes

Partial progress on #1454 (Phase 3c)

[github-actions]

Rule Review: Clean refactor with one minor docstring inaccuracy

Rules checked: operations.md, code-style.md, testing.md, git-workflow.md
Files reviewed: 7 (.claude/rules/operations.md, node.rs, p2p_protoc.rs, op_state_manager.rs, get/op_ctx_task.rs, put/op_ctx_task.rs, subscribe.rs)

---

Warnings

None.

---

Info

• crates/core/src/node/op_state_manager.rs:196–199 — The new docstring on SubOperationTracker::is_sub_operation reads "This method remains for legacy relay paths and unit tests" but the method is #[cfg(test)], making it inaccessible to production relay-path code. The phrasing implies production use that can't exist. Should read "This method remains for unit tests of legacy relay paths" (code-style.md: doc comments should accurately describe what the code does).

---

Summary of changes reviewed: The PR correctly replaces four op_manager.is_sub_operation() call-sites (DashMap lookup via SubOperationTracker) with Transaction::is_sub_operation() (structural parent.is_some() check). The structural check is a strict superset — it works for both legacy and task-per-tx children, including the latent bug class of async non-blocking subscribe children that were never registered in the tracker. The expect_and_register_sub_operation removals from PUT and GET task-per-tx drivers are correct: those drivers never used the parking-lot mechanism that registration was designed to feed. The operations.md update and line-number reference (operations.rs:182–208) are accurate against the current file. The removed test assertions (ordering of expect_and_register_sub_operation) are correctly dropped since the call they guarded no longer exists.

---

Rule review against .claude/rules/. WARNING findings block merge.