fix(radio): cache mod params on SET_CONFIG so TX works without prior RX

The TX state machine's `prep_for_cad_act` (TxCadPrep entry) and
`prep_for_tx_act` (TxPrep entry) both read `ctx.mdltn` and emit
`TransmitFailed` → `ERR(ERadio)` if it's None. But `ctx.mdltn` was
only ever assigned on the RX path — `try_start_rx_hw` (RxArming entry)
and `cache_rx_pkt` (RxSession entry). `reconfigure_radio` already
called `to_lora_mod_params(lora, cfg)` during SET_CONFIG but discarded
the result via `let _ = …?;` (comment claimed image calibration; the
function is pure compute on SX126x).

Result: a freshly-booted dongle that issues TX before ever calling
RX_START gets `ERR(ERadio)` on the first CmdTx because the chart's
TX-entry actions find `ctx.mdltn = None`. The bug was hidden in
practice by any prior RX_START (ai-bot, ping_pong --role rx, etc.)
leaving `ctx.mdltn` populated across sessions until the next reflash.

Fix: thread the `ModulationParams` out of `reconfigure_radio` as its
return value, and have `apply_set_config` cache it in `self.mdltn`
alongside `self.config`. The failure path clears `self.mdltn` along
with `self.config` so a SET_CONFIG that fails leaves a clean
Unconfigured state.

Surfaced while testing the matching sync-word fix on freshly-flashed
dongles via `examples/sync_word_test.py` — the first TX failed with
`ERR(ERadio)` until this path was wired up.

src/radio.rs

@@ -1252,7 +1252,9 @@ impl RadioActions for RadioActionContext<'_> {
                     Some(lora) => reconfigure_radio(lora, &cfg).await,
                     None => Err(RadioError::Busy),
                 };
-                if recfg_result.is_err() {
+                let mdltn = match recfg_result {
+                    Ok(m) => m,
+                    Err(_) => {
                         // Hardware failure: drop to Unconfigured (spec §6.3).
                         // Clear all per-session state — including any in-flight
                         // RxStart tag — so the next CmdRxStart in Unconfigured
@@ -1274,7 +1276,9 @@ impl RadioActions for RadioActionContext<'_> {
                         }
                         return;
                     }
+                };
                 self.config = Some(cfg);
+                self.mdltn = Some(mdltn);
                 info!("config set (apply_set_config success tag={=u16})", tag);
                 self.cad_retry_delay_us = toa_us(&cfg, CAD_ASSUMED_PEER_PAYLOAD_B);
                 if self.cad_retry_delay_us > CAD_BUDGET_US {
@@ -2094,7 +2098,10 @@ fn to_lora_mod_params(
     lora.create_modulation_params(sf, bw, cr, cfg.freq_hz)
 }
 
-async fn reconfigure_radio(lora: &mut LoRaDriver, cfg: &LoRaConfig) -> Result<(), RadioError> {
+async fn reconfigure_radio(
+    lora: &mut LoRaDriver,
+    cfg: &LoRaConfig,
+) -> Result<lora_phy::mod_params::ModulationParams, RadioError> {
     lora_setup("reconfigure_radio.enter_standby", lora.enter_standby()).await?;
     // Apply the host-supplied LoRa sync word. The protocol field is a
     // u16 holding the SX126x register pair in big-endian order
@@ -2112,10 +2119,7 @@ async fn reconfigure_radio(lora: &mut LoRaDriver, cfg: &LoRaConfig) -> Result<()
         lora.set_lora_sync_word(sync_byte),
     )
     .await?;
-    // Calibrate image etc. via a throwaway create_modulation_params call
-    // — inside the SET_CONFIG OK window so first TX/RX is fast.
-    let _ = to_lora_mod_params(lora, cfg)?;
-    Ok(())
+    to_lora_mod_params(lora, cfg)
 }
 
 fn validate_lora(cfg: &LoRaConfig, info: &Info) -> Result<(), ErrorCode> {