> ## Documentation Index
> Fetch the complete documentation index at: https://whatsapp-rust.jlucaso.com/llms.txt
> Use this file to discover all available pages before exploring further.

# June 10, 2026 — Group send: server-aware LID-PN probe on warm device lookups

> get_devices_from_registry now probes only the single relevant direction of the LID-PN cache per member, cutting device resolution from 3 cache operations to 2 and reducing warm 800-member send latency by 26%.

## Performance

**Group send: directed LID-PN probe in `get_devices_from_registry` ([#823](https://github.com/oxidezap/whatsapp-rust/pull/823))**

Each call to `get_devices_from_registry` resolves a `Jid` to its canonical lookup keys before hitting the device registry cache. Previously, the key resolver (`resolve_lookup_keys`) only received the bare user string. It blindly probed **both** directions of the LID-PN map — `get_phone_number` then `get_current_lid` — for every group member. This resulted in 3 moka lookups per participant before the registry lookup.

**Server-aware probe.** A hot caller in `get_devices_from_registry` already holds the full `Jid`, and the namespace fully determines which direction is meaningful:

* A **LID** user (`@lid`) can only appear in the `lid → pn` map.
* A **PN** user (`@s.whatsapp.net`) can only appear in the `pn → lid` map.

The new `resolve_lookup_keys_for_jid` dispatches on the JID's server field and probes exactly one direction for LID/PN JIDs. All other namespaces keep the existing two-probe fallback. This turns 3 cache operations per member into 2 with no change in the resulting canonical keys.

> A concurrent-resolution prototype (`buffered(50)` stream, mirroring WA Web's `SESSION_CHECK` batching) was measured and rejected: 962 µs vs. 675 µs serial for a warm 800-member send (+43% slower). Moka hits resolve immediately, so there is no I/O to overlap and the stream machinery is pure overhead. The reliable gain comes from doing less work per member, not reordering the same work.

**Benchmark** (release, warm caches, 800 members × 2 devices, 500 iterations):

|                            | Warm 800-member send |
| -------------------------- | -------------------- |
| Before (blind two-probe)   | 675 µs               |
| After (server-aware probe) | 497 µs (**−26%**)    |

**No breaking changes.** `get_devices_from_registry` keeps its existing signature; the directed probe is an internal implementation detail.