fix(snr): review findings 3/5 — live getStats overlay (no meter freeze during bursts) + session-teardown SNR clear; finding 7 filed (TNC wire semantics need the cross-machine rig)

Verified: ctest 85/85; good@20 gate PASS 1530 bps.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019btwFxDt7D19SZSqPYvZnk

Author: secup

docs/CHANGELOG.md

@@ -10,6 +10,37 @@ This log tracks all bug fixes and behavioral changes to prevent re-doing work du
 
 ---
 
+## 2026-07-10 — fix(snr): review findings 3/5 — getStats() live overlay (meters no longer freeze during bursts) + session-teardown SNR clear; finding 2 script removed; finding 7 filed
+
+**Finding 3:** the LoopbackStats cache is refreshed only by the per-frame
+callback, which the burst transport suppresses — GUI meters (and any polled
+consumer of getStats) froze at handshake-era values for entire file transfers
+while the protocol correctly consumed the decoder's lock-free atomics.
+getStats() now OVERLAYS those live values (per-frame OFDM broadband — updated
+inside bursts — plus the live idle-noise meter) so one snapshot serves all
+consumers. The broadband atomics are already cleared by decoder reset, so the
+overlay cannot resurrect a dead session's values.
+
+**Finding 5:** disconnect never cleared the stats cache — the previous QSO's
+SNR/fading stayed on the meters and could be attached to the next session's
+first PONG. ModemEngine::setConnected(false) now clears the channel-
+measurement fields; the live idle overlay repopulates within ~1 s.
+
+**Finding 2:** tools/snr_meter_validation.sh REMOVED (REMOVAL_BACKLOG R12) —
+it false-passed via a deleted probe's stale binary + column mismatch + empty
+summary treated as success; superseded by the OFDMSnrCalibration CTest matrix.
+
+**Finding 7 (FILED, not changed):** the TNC's polled SNR reads the protocol
+engine's measured value, which defaults to 15 dB (source NONE) before any
+measurement, and conflates unknown with 0 dB. Changing what ultra_tnc reports
+to client apps is a wire-behavior change we won't ship without the
+cross-machine Winlink/PAT rig to test against — filed in the SNR handoff §7
+follow-ups with this analysis.
+
+Verified: ctest 85/85; good@20 gate PASS 1530 bps.
+
+---
+
 ## 2026-07-10 — fix(snr): external-review findings 1/4/6 — in-band-first noise source (guard bins lie on band-limited noise), physical-SNR per-frame validity, session-only ring clear
 
 **Finding 1 (critical, real-radio class):** the OFDM meter PREFERRED guard-bin

docs/SNR_CALIBRATION_HANDOFF_2026_07_08.md

@@ -243,6 +243,18 @@ handed over: MPG@0 connect + 5 frames (ack path dead); MPG@5 connect + ~98%
 (endgame ack); MPG@10 repeated full CRC deliveries. Graded degradation, every
 failure named.
 
+## §9 (external review 2026-07-10): TNC SNR reporting semantics — FILED
+
+The TNC's polled SNR reads ProtocolEngine::getMeasuredSNR (Connection default
+15.0 dB, source NONE) — before any measurement a client sees a confident "15";
+unknown and a genuine 0 dB are indistinguishable. Fix shape: report the
+modem's live getStats() overlay when the protocol source is NONE, and give the
+TNC status string an explicit unknown marker — BUT this changes what client
+apps parse, so it needs the cross-machine Winlink/PAT rig for validation
+(memory: project_winlink_pat_otasim_crossmachine, tnc bulk-accum notes)
+before shipping. Everything else from the 2026-07-10 external review
+(findings 1-6) is fixed and gated — see CHANGELOG entries of that date.
+
 ## Traps that cost hours tonight (do not rediscover)
 
 1. **IONOS is an S:N machine** — noise level TRACKS the input signal. Idle

src/gui/modem/modem_engine.cpp

@@ -941,8 +941,36 @@ Bytes ModemEngine::getReceivedData() {
 }
 
 LoopbackStats ModemEngine::getStats() const {
-    std::lock_guard<std::mutex> lock(stats_mutex_);
-    return stats_;
+    LoopbackStats out;
+    {
+        std::lock_guard<std::mutex> lock(stats_mutex_);
+        out = stats_;
+    }
+    // LIVE OVERLAY (external review 2026-07-10, finding 3): the cache above is
+    // refreshed only by the per-frame callback, which the burst transport
+    // suppresses (frames deliver as a unit) — so during a file transfer every
+    // getStats() consumer (GUI meters, polled TNC) froze at the handshake-era
+    // values while the protocol correctly consumed the decoder's lock-free
+    // atomics. Overlay those same live values here so one snapshot serves all
+    // consumers: per-frame OFDM broadband (updated inside bursts), the live
+    // idle-noise meter, and the current fading index.
+    if (streaming_decoder_) {
+        if (streaming_decoder_->hasLastOFDMBroadbandSNREstimate()) {
+            const float bb = streaming_decoder_->getLastOFDMBroadbandSNREstimate();
+            if (std::isfinite(bb)) {
+                out.snr_db = bb;
+                out.snr_source = SNRSource::OFDM_BROADBAND;
+                out.has_ofdm_broadband_snr_db = true;
+                out.ofdm_broadband_snr_db = bb;
+            }
+        }
+        const auto idle = streaming_decoder_->getIdleNoiseSNRSnapshot();
+        if (idle.valid && std::isfinite(idle.idle_in_band_snr_db)) {
+            out.has_idle_in_band_snr_db = true;
+            out.idle_in_band_snr_db = idle.idle_in_band_snr_db;
+        }
+    }
+    return out;
 }
 
 DecoderStats ModemEngine::getDecoderStats() const {

src/gui/modem/modem_mode.cpp

@@ -194,6 +194,22 @@ void ModemEngine::setConnected(bool connected) {
         // Switching to disconnected state - use robust mode for RX
         decoder_->setRate(CodeRate::R1_4);
 
+        // SESSION SNR LIFECYCLE (external review 2026-07-10, finding 5): the
+        // LoopbackStats cache is refreshed only by decode events, so without
+        // this the previous QSO's SNR/fading stayed visible on the meters and
+        // could be attached to the next session's first PONG. Clear the
+        // channel-measurement fields at teardown; the idle meter (live overlay
+        // in getStats) repopulates within ~1 s of idle audio.
+        updateStats([&](LoopbackStats& st) {
+            st.snr_db = 0.0f;
+            st.snr_source = SNRSource::NONE;
+            st.has_ofdm_broadband_snr_db = false;
+            st.ofdm_broadband_snr_db = 0.0f;
+            st.ofdm_internal_snr_db = 0.0f;
+            st.sync_quality_db = 0.0f;
+            st.synced = false;
+        });
+
         // BUG-TNC-SESSION-001 fix (port from tools/ultra_tnc.cpp): a persistent
         // ModemEngine across multiple sessions must perform the same RX/TX
         // state reset on disconnect that ultra_tnc does, otherwise the next