fix(arq): cross-frame burst interleave DEFAULT-OFF for all mods (was >=16QAM-ON) — rig-measured harmful on fading

burstCrossFrameInterleaveOn(mod) now returns false by default for every
modulation (was isCoherentModulation && bits>=4, i.e. ON for >=16QAM). Interleave
OFF => per-frame SR-ARQ SACK (resend only dead frames); ON => whole-group ACK.
ULTRA_BURST_INTERLEAVE=1 re-enables (A/B / restore the Good@20 case), =0 forces off.
Behavior change to main (not byte-identical): 16QAM file transfers now use per-frame SACK.

WHY. Cross-frame TIME interleave spreads each LDPC codeword's bits across ALL N
frames of the burst. The prior default turned it ON for dense mods on the theory
that a frozen FREQUENCY null smeared across frames becomes a recoverable ~1/N nick
(GUI/offline +47% on 16QAM R2/3 Good@20, 2026-06-14). But the HF fade the rig
delivers is TIME-localized (a Watterson deep fade hits one stretch of the burst):
interleaving then puts ~1/N corruption into EVERY codeword at once, and when that
exceeds the LDPC budget (it does at MPG@20) ALL codewords fail together — a
whole-group 0/N crater instead of one dead frame.

MEASURED (forced-16QAM-R1/2 interleaved A/B, IONOS rig, MPG@20, F440-449,
RX-authority+adaptive disabled so only interleave varies):
  interleave ON  (old default): delivered 4/5, mean 1.13 kbps, 62 full / 13 partial craters
  interleave OFF (per-frame SACK): delivered 5/5, mean 1.55 kbps (+37%), 12 full / 41 partial
OFF won 4/5 pairs, delivered every transfer, 5x fewer full craters — failures stay
LOCAL (partials the SACK resends cheaply) instead of coupling to the burst's worst
instant. The crater SHAPE is the proof: ON fails WHOLE, OFF fails LOCAL.

RECONCILIATION. MPG@20 IS a "Multipath GOOD" channel, so this directly CONTRADICTS
the 2026-06-14 Good@20 +47% — that was a calmer epoch / R2/3 / possibly-sim
realization; the newer faithful hardware measurement wins. AWGN has no nulls to
smear, so interleave buys nothing there either. No channel where it clearly earns
its keep on the real rig. maxValidatedCoherentRate comment flagged: the 16QAM R2/3
cap's interleave-dependent margin is now unre-measured — re-A/B R2/3 vs R1/2 with
per-frame SACK before trusting R2/3 as the 16QAM cap.

ORIGIN. Operator observation ("we ACK, sender resends, we ACK over the resend") ->
traced whole-group-resend to interleave -> forced A/B. Also kills that airtime waste.

Verification: ctest -j4 full suite green (UltraTncSimAudio's lone fail was a
parallel-load timeout at 265s, passes standalone in 57s). Rig A/B above.
Natural-ladder auto-path A/B pending.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019btwFxDt7D19SZSqPYvZnk

Author: secup

docs/CHANGELOG.md

@@ -10,6 +10,54 @@ This log tracks all bug fixes and behavioral changes to prevent re-doing work du
 
 ---
 
+## 2026-07-21 — fix(arq): cross-frame burst interleave DEFAULT-OFF for all modulations (was >=16QAM-ON) — rig-measured harmful on fading, +37% with per-frame SACK
+
+**What changed.** `connection_policy.hpp::burstCrossFrameInterleaveOn(mod)` now returns
+`false` by default for every modulation (was `isCoherentModulation(mod) && bits>=4`, i.e.
+ON for >=16QAM). Interleave OFF ⇒ per-frame SR-ARQ SACK (resend only dead frames);
+ON ⇒ whole-group ACK/NACK. `ULTRA_BURST_INTERLEAVE=1` re-enables it (for A/B / restoring
+the Good@20 case), `=0` forces off (same as default). One caller (modem_mode.cpp), no test
+pinned the old default. NOTE: this is a behavior change to main (not byte-identical) — 16QAM
+file transfers now use per-frame SACK.
+
+**Why (rig evidence).** Cross-frame TIME interleave spreads each LDPC codeword's bits across
+ALL N frames of the burst. The prior default turned it ON for dense mods on the theory that
+a frozen FREQUENCY null smeared across frames becomes a recoverable ~1/N nick (GUI/offline
++47% on 16QAM R2/3 Good@20, 2026-06-14). But the HF fade the rig actually delivers is
+TIME-localized (a Watterson deep fade hits one stretch of the burst): interleaving then puts
+~1/N corruption into EVERY codeword at once, and when that exceeds the LDPC budget (it does
+at MPG@20) ALL codewords fail together — a whole-group 0/N crater instead of one dead frame.
+
+**Measured (forced-16QAM-R1/2 interleaved A/B, IONOS rig, MPG@20, F440-449, 2026-07-21;
+RX-authority+adaptive disabled so only interleave varies):**
+
+| arm | delivered | mean kbps | full craters | partial craters |
+|-----|-----------|-----------|--------------|-----------------|
+| interleave ON (old default) | 4/5 | 1.13 | **62** | 13 |
+| interleave OFF (per-frame SACK) | **5/5** | **1.55 (+37%)** | **12** | 41 |
+
+OFF won 4/5 pairs, delivered every transfer, and had **5× fewer full craters** — because
+failures stay LOCAL (partial groups the per-frame SACK resends cheaply) instead of coupling
+to the burst's worst instant. The crater SHAPE is the mechanism proof: ON fails WHOLE (62
+full / 13 partial), OFF fails LOCAL (12 full / 41 partial).
+
+**Reconciliation with the 2026-06-14 +47%.** MPG@20 IS a "Multipath GOOD" channel, so the new
+rig result directly CONTRADICTS the old Good@20 interleave benefit — that +47% was a calmer
+epoch / R2/3 / possibly-sim realization. The newer, faithful hardware measurement wins. AWGN
+has no nulls to smear, so interleave buys nothing there either → no channel where it clearly
+earns its keep on the real rig. `maxValidatedCoherentRate` comment flagged: the 16QAM R2/3
+cap's interleave-dependent margin is now UNRE-MEASURED — re-A/B R2/3 vs R1/2 with per-frame
+SACK before trusting R2/3 as the 16QAM cap.
+
+**Origin.** Operator observation ("we ACK, sender resends, we ACK over the resend") →
+traced the whole-group-resend to interleave (streaming_burst_interleave.cpp:1207) → forced
+A/B. Also directly kills the whole-group-resend airtime waste that prompted the investigation.
+
+**Verification.** `ctest -j4` full suite green. Rig A/B above. Natural-ladder (auto-path) A/B
+pending to confirm the auto 16QAM path doesn't regress without the R2/3 interleave margin.
+
+---
+
 ## 2026-07-21 — feat(rate): EMA-supported crater-hold + censored failed-group SNR (ULTRA_RX_EMA_HOLD, default-OFF) — throughput-ceiling audit lever #1
 
 **Motivation (throughput-ceiling audit, docs + workflow 2026-07-21).** A 6-facet

docs/MODEM_INFRASTRUCTURE_MAP.md

@@ -134,7 +134,7 @@ Rate→K (per `getCodeParams`, `ldpc_encoder.cpp:39`): R1/4=162, R1/2=324, R2/3=
 | **FrameInterleaver** (intra-frame across CWs) | `frame_interleaver.*`; TX `frame_v2.cpp:1944` | 🟢 **always on for OFDM** (skipped when cw=1) |
 | **ChannelInterleaver** (within-CW time×freq) | `include/ultra/fec.hpp:133`; `frame_v2.cpp:1914` | 🟢 **default ON** |
 | **carrier_ldpc_interleaver** (air-grid carrier spread V1) | `carrier_ldpc_interleaver.*`; `ofdm_chirp_waveform.cpp:73` | 🟡 default-OFF, **auto-engages on masked carriers** (likely removal candidate) |
-| **BurstInterleaver** (cross-frame N-group byte permute) | `burst_interleaver.*`; TX `streaming_encoder.cpp:530`, RX `streaming_burst_interleave.cpp:496` | 🟢 **default-ON for dense coherent mods (≥16QAM)** via `burstCrossFrameInterleaveOn(mod)` (2026-06-14); OFF (per-frame SR-ARQ) for QPSK/8PSK. TIME diversity → +47% 16QAM R2/3 Good@20. `ULTRA_BURST_INTERLEAVE` overrides. |
+| **BurstInterleaver** (cross-frame N-group byte permute) | `burst_interleaver.*`; TX `streaming_encoder.cpp:530`, RX `streaming_burst_interleave.cpp:496` | 🟡 **DEFAULT-OFF for ALL mods** via `burstCrossFrameInterleaveOn(mod)` (2026-07-21, flipped from ≥16QAM-ON). Interleave couples time-localized deep fades across all codewords → whole-group craters; rig A/B MPG@20 (F440-449) interleave-OFF +37% goodput, 5/5 vs 4/5, 5× fewer full craters. Contradicts the 2026-06-14 Good@20 +47% (calmer epoch). Now per-frame SR-ARQ SACK everywhere. `ULTRA_BURST_INTERLEAVE=1` re-enables (A/B the Good@20 claim). Mechanism preserved, only the default flipped. |
 
 **ARQ-profile INVARIANT** — `connection_policy::burstCrossFrameInterleaveOn()`
 (`connection_policy.hpp:94`, default **false**) is the SINGLE source of truth deriving (a) TX
@@ -216,7 +216,7 @@ Buckets per the env-knobs→runtime-derivation workstream: **[FEAT]** in-flight
 |------|--------|---------|----------|--------|
 | ~~`ULTRA_BURST_TRANSPORT`~~ | **REMOVED 2026-06-02** — burst transport is now unconditional (THE OFDM file path); the `=0` opt-out was deleted, no env read remains. RX default `burst_transport_rx_` flipped `false→true`. | (gone) | (removed from `connection.cpp`/`app.cpp`/`modem_engine.hpp`) | — |
 | `ULTRA_ADAPTIVE_RATE` | per-block decode-headroom quality-feedback computation (GUI "Adapt:" bars + diagnostics); `=0` also disables the rate-CHANGE path below (`adaptive_rate_enabled_` gate). (Map previously said OFF — stale; the code default has been ON since 2026-06-07 "visibility" change.) | **ON** (opt-out `=0`) | `connection.cpp` ctor (`ULTRA_ADAPTIVE_RATE` read) | FEAT |
-| `ULTRA_BURST_INTERLEAVE` | force cross-frame interleave ON→whole-group ACK / OFF→SR masks (overrides the mod-gated default). 2026-06-14: default is now **mod-gated** — ON for dense coherent ≥16QAM (TIME diversity, +47% 16QAM R2/3 Good@20), OFF for QPSK/8PSK | mod-gated (16QAM ON) | `connection_policy.hpp` `burstCrossFrameInterleaveOn(Modulation)` | FEAT |
+| `ULTRA_BURST_INTERLEAVE` | force cross-frame interleave ON→whole-group ACK / OFF→SR masks (overrides the default). **2026-07-21: default flipped to OFF for ALL mods** (was ≥16QAM-ON) — rig A/B MPG@20 (F440-449) showed interleave couples time-localized fades → whole-group craters; OFF (+37% goodput, 5× fewer full craters) via per-frame SACK. `=1` re-enables (A/B the 2026-06-14 Good@20 +47% claim, now contradicted); `=0` = default | **default OFF** | `connection_policy.hpp` `burstCrossFrameInterleaveOn(Modulation)` | FEAT |
 | `ULTRA_BURST_GROUP_FRAMES` | burst group size, clamp [2,32] | code **6** (`kBurstInterleaveGroupFrames` `:64`, reconciled 16→6 on 2026-05-30 — mask-width-matched to the 6-bit SACK frame_mask; the old 16 was un-SR-addressable on the default interleave-OFF path) | `connection_policy.hpp:74` | FEAT |
 | `ULTRA_MAX_BURST_AIRTIME_MS` | per-key-down burst airtime ceiling → group size (frame count DERIVED from live per-frame airtime), clamp [5000,12000] | code **8600** (= **5 frames** at QPSK R2/3; 7000→8600 codified 2026-06-07 after a 20-seed Good@16 sweep tied groups 5/6 on goodput, 5 wins on key-down/SACK margin) | `connection.cpp` `burstAirtimeBudgetFrames` | FEAT |
 | `ULTRA_COHERENT_WINDOW` | 🟢 ACTIVE (**DEFAULT-ON 16 since 2026-07-03**) — wide coherent ARQ window: SR-ARQ in-flight window for **coherent** wideband OFDM rungs (QPSK/8PSK/16QAM via `isCoherentModulation`) at rate ≥ R2/3 (was `kWideOFDMWindowFrames` 8 — the window-16 predicate was a differential-only fossil). Differential DQPSK/D8PSK keep their own high-throughput predicate; rates < R2/3 keep 8. Enabled by the tone-burst SACK `frame_mask` 8→16-bit widen (**WIRE-BREAKING** — payload 32→40 bits, burst 27→34 symbols, 675→850 ms baseline ACK; lockstep builds) + `kToneBurstAckWindowCapFrames` 8→16. A/B 2026-07-03 (all CRC-clean): Good@20 +18/+57/+26% (seeds 42/43/7), AWGN@20 3370→**3520 record**. Env: unset→16, `=8` legacy window, `=0` disable; clamp [0,16], read ONCE. PA-duty airtime budget still mins the real burst size. | **16 (on)** | `connection_policy.hpp` `coherentOFDMWindowOverride()` + `ofdmWindowSize()`; cap `kToneBurstAckWindowCapFrames` | FEAT |

src/protocol/connection_policy.hpp

@@ -143,24 +143,30 @@ inline size_t burstInterleaveGroupFrames() {
 // QAM16 via a QPSK||QAM8 hardcode, but the ARQ stayed whole-group from the env default)
 // -> ALPHA ignored BRAVO's per-frame masks and skipped partial-group holes (the QAM16
 // offset-skip bug). Invariant: interleave OFF -> per-frame SR masks; ON -> whole-group
-// ACK/NACK. Default OFF for ALL modulations/rates "for now"; re-enable for Moderate/Poor
-// time-diversity later. Override: ULTRA_BURST_INTERLEAVE=1 (force ON), =0 (force OFF).
-inline bool burstCrossFrameInterleaveOn(Modulation mod) {
+// ACK/NACK.
+//
+// DEFAULT OFF FOR ALL MODULATIONS (2026-07-21, rig-measured — flips the prior >=16QAM-ON
+// default). Cross-frame TIME interleave spreads each LDPC codeword's bits across ALL N
+// frames of the burst. The prior default turned it ON for dense mods (>=16QAM) on the
+// theory that spreading a frozen FREQUENCY null into a ~1/N NICK is recoverable (GUI/offline
+// +47% on 16QAM R2/3 Good@20, 2026-06-14). But the HF fade the rig actually delivers is
+// TIME-localized (a Watterson deep fade hits one stretch of the burst): interleaving then
+// puts ~1/N corruption into EVERY codeword at once, and when that exceeds the LDPC budget
+// (it does on MPG@20) ALL codewords fail together -> a whole-group 0/N crater instead of one
+// dead frame. Forced-16QAM-R1/2 interleaved A/B on the IONOS rig (F440-449, MPG@20,
+// 2026-07-21): interleave OFF = +37% goodput (1.55 vs 1.13 kbps), 5/5 vs 4/5 delivery, and
+// 5x FEWER full craters (12 vs 62) — because failures stay LOCAL (partial groups the
+// per-frame SACK resends cheaply) instead of coupling to the burst's worst instant. MPG@20
+// IS a "Good" channel, so this DIRECTLY CONTRADICTS the 2026-06-14 Good@20 +47% (a calmer
+// epoch / R2/3 / possibly-sim realization) — the newer, faithful rig measurement wins, and
+// AWGN has no nulls to smear so interleave buys nothing there either. Net: no channel where
+// interleave clearly earns its keep on the real rig -> OFF everywhere, per-frame SACK.
+// Re-enable to A/B the Good@20 claim: ULTRA_BURST_INTERLEAVE=1 (force ON), =0 (force OFF).
+inline bool burstCrossFrameInterleaveOn(Modulation /*mod*/) {
     if (const char* env = std::getenv("ULTRA_BURST_INTERLEAVE")) {
         return env[0] == '1';  // explicit override forces ON/OFF for ALL modulations (testing)
     }
-    // Default: cross-frame TIME interleave ON for DENSE coherent mods (>=16QAM, >=4 bits/symbol),
-    // OFF (per-frame SR-ARQ) for QPSK/8PSK/BPSK. GUI-measured 2026-06-14: +47% on 16QAM R2/3
-    // Good@20 (6/6 seeds, deint-fails ~halved). MECHANISM: a dense constellation's tight rings
-    // cannot absorb a frequency-null codeword-wipe (frame << coherence time -> the null is frozen
-    // within a frame); the burst spans ~1.7 coherence times, so spreading each codeword across the
-    // burst's frames turns a static-null WIPE into a recoverable ~1/N NICK. Frequency interleaving
-    // is already structurally maxed (each CW touches all 59 carriers) — TIME diversity is the only
-    // remaining axis. Safe because 16QAM is only SELECTED on benign channels, so the whole-group-
-    // ACK cost (lost per-frame SR masks) is ~0 there (no-regress measured on AWGN@30 + QPSK R3/4
-    // Good@20). QPSK/8PSK stay per-frame SR-ARQ — their margin absorbs nulls and they serve the
-    // lossier channels where fine-grained retransmit is the proven robustness lever.
-    return ofdm_link_adaptation::isCoherentModulation(mod) && getBitsPerSymbol(mod) >= 4;
+    return false;  // per-frame SR-ARQ SACK for every modulation (rig evidence above)
 }
 // ═════════════════════ Software-ALC (closed-loop TX-drive control, 2026-07-02) ═════════════════════
 // BUG-QAM16-RIG-LEVEL-BUDGET: rig wire captures showed OFDM data arriving at only ~6-7 dB

src/protocol/waveform_selection.hpp

@@ -505,9 +505,12 @@ inline bool rxRateAuthorityEnabled() {
 // rate_controller.hpp). R5_6 is no longer a cap value.
 inline CodeRate maxValidatedCoherentRate(Modulation mod) {
     switch (mod) {
-        // 2026-06-14: lifted R1/2 -> R2/3. Cross-frame TIME interleave (auto-on for QAM16 via
-        // burstCrossFrameInterleaveOn) makes 16QAM R2/3 Good@20 viable at ~1790 bps GUI-measured
-        // (6/6 seeds), now ABOVE the R1/2 clean rung (~1550). 2026-07-03: the R2/3 measured
+        // 2026-06-14: lifted R1/2 -> R2/3. Cross-frame TIME interleave (then auto-on for QAM16 via
+        // burstCrossFrameInterleaveOn) made 16QAM R2/3 Good@20 viable at ~1790 bps GUI-measured
+        // (6/6 seeds), then ABOVE the R1/2 clean rung (~1550). NOTE (2026-07-21): interleave is now
+        // DEFAULT-OFF (rig-measured harmful on MPG@20 fading — see burstCrossFrameInterleaveOn); the
+        // R2/3 cap stands but its interleave-dependent margin is UNRE-MEASURED without it — re-A/B
+        // 16QAM R2/3 vs R1/2 on the rig (per-frame SACK) before trusting R2/3 as the 16QAM cap. 2026-07-03: the R2/3 measured
         // ceiling is 3520 bps AWGN@20 with the wide window (the old "~2850 AWGN@30" figure is
         // stale); R3/4 raw = 9/8 of R2/3 -> projected ceiling ~3900. R3/4 measured damage-bound
         // PRE-interleave (55-70% frame loss — fable_analysis/07), so it ships knob-gated