cameri / nostream / 41332816237

78%

main: 78%

push

github

refactor(relay): decode the refusal word exhaustively (#1026)

Closes #993.

`Account::refusal()` decoded the refusal word with a wildcard:

```rust
match self.refusal.load(Ordering::SeqCst) {
    QUEUE_FULL => Some(Refusal::QueueFull),
    BUDGET_EXHAUSTED => Some(Refusal::BudgetExhausted),
    _ => None,           // an unknown word reads as "no refusal"
}
```

`Refusal` has two variants, so the wildcard is unreachable *today*. The
defect is what a third cause does: `refuse()` and `close_refused()` are
compiler-forced to handle it, `refusal()` is not, and every reader takes
its `None` as "nothing to do".

**Measured, not argued.** A program carrying the decode verbatim, over
four words:

```
       NO_REFUSAL (0) -> None; is_some()=false
       QUEUE_FULL (1) -> Some(QueueFull); is_some()=true
 BUDGET_EXHAUSTED (2) -> Some(BudgetExhausted); is_some()=true
    A_THIRD_CAUSE (3) -> None; is_some()=false     <-- indistinguishable from no refusal
```

while the same snippet with a third variant added does not compile at
the encode site (`error[E0004]: non-exhaustive patterns`). That
asymmetry is the issue.

## What changed

The refusal word becomes a `Refusal` that names its own "nothing
refused" case:

```rust
#[repr(u8)]
pub enum Refusal {
    None = NO_REFUSAL,              // 0 — the value the first-cause-wins CAS compares
    QueueFull = QUEUE_FULL,         // 1
    BudgetExhausted = BUDGET_EXHAUSTED, // 2
}
```

`Account::refusal()` and `FrameSink::refusal()` return `Refusal` rather
than `Option<Refusal>`; `word()`/`from_word()` both live on the type;
and every reader that decides a cause now has to answer the compiler:

| site | before | after |
|---|---|---|
| `sink.rs` `send`, two latch checks | `== Some(Refusal::QueueFull)` |
`== Refusal::QueueFull` (still an equality test — see below) |
| `relay.rs` connection loop | `if let Some(refusal) = …` | `match` with
`Refusal::None => {}` and `refusal @ (QueueFull \| BudgetExhausted)` |
| `relay.rs` `... (continued)

3588 of 5362 branches covered (66.92%)

Branch coverage included in aggregate %.

7992 of 10291 relevant lines covered (77.66%)

867.55 hits per line