Epic: spatial-diversity axis — exploiting idle RF chains for the single-stream video link

[josephnef]

The third diversity axis

devourer's long-range single-stream video link already exploits two physical-layer
diversity axes:

• frequency — single-adapter per-packet hopping (docs/frequency-hopping.md),
  which doubles as a frequency-diversity interleaver for the outer FEC;
• code — the fused cross-layer FEC stack (docs/fused-fec.md): sub-block
  integrity salvage, an MDS outer code, soft-decision inner decode, and
  per-temporal-layer unequal error protection.

The spatial axis is unexploited. Because a single-stream video link never
stacks spatial streams, the extra RF chains (1 idle on a 2T2R part, up to 3 on a
4T4R 8814) sit unused. Turning them into diversity/robustness rather than
throughput is almost pure upside for this use case: it is USB-free (one
stream in, one stream out — no MIMO-multiplexing bandwidth hit), and it composes
multiplicatively with the frequency and code axes already in place.

The single hard constraint (from docs/frequency-hopping.md): there is one LO
per chip, so the chains share one centre frequency — they give spatial
diversity within a channel, orthogonal to (not a substitute for) the frequency
diversity that hopping provides.

Why this matters

• RX MRC across a 4T4R chip's chains is a ~10·log₁₀(N) dB SNR win from a
  single ground adapter — partially replacing wfb-ng's multi-adapter diversity
  array, in-chip, at zero USB cost.
• TX diversity (CSD / STBC) attacks the deep fades that push corruption
  into the frame-wide regime where sub-block salvage fails — so spatial diversity
  improves the precondition of the whole fused-FEC stack, not just its input.
• Chain count is a new lever for the energy-minimizing adaptive link
  (docs/adaptive-link.md), which today has no spatial knob: a convex
  PA-power-saved vs chain-baseline-added tradeoff with an optimal N per range,
  and a per-layer spatial UEP knob alongside modulation and FEC.

Approach

Research/design track — simulators, measurement harnesses, and conceptual docs
first (the way the frequency and code axes were developed), hardware validation
following. Each activity is a separate tracked issue.

Roadmap

Foundations

• HAL capability audit — independent per-chain TX (CSD/STBC/beam-steer vs RF-path enable) #128 — HAL capability audit ✅ (capability matrix posted)
• Antenna-decorrelation measurement harness #129 — Antenna-decorrelation measurement harness ✅ (merged PR Add antenna-decorrelation measurement harness (spatial-diversity #129) #140; N_eff metric)

Core diversity techniques

• RX-side in-chip MRC on an 8814 ground station #130 — RX-side in-chip MRC ✅ measured on hardware (features merged PR RX-chain masking + CCK TX rates (spatial-diversity #130) #141)
• ANTSEL per-frame antenna-selection diversity probe (rescoped from CSD) #131 — ANTSEL selection ✅ closed (superseded: RX-MRC dominates it; power→Spatial × energy — active-chain count as an adaptive-link lever #135, TX→Close STBC on-air — resolve the 2-channel-alignment root cause #132)
• Close STBC on-air — resolve the 2-channel-alignment root cause #132 — STBC on-air ✅ (PR Chip STBC TX diversity + txdemo event-loop EINTR fix (#132) #144): chip emits STBC (6336/6336), and TX-STBC diversity MEASURED — 1.94× static → 3.27× moving at marginal MCS5 (TX-side dual of RX-side in-chip MRC on an 8814 ground station #130, helps even a single-antenna RX)

Cross-layer composition

• Spatial × FEC — spatial diversity keeps corruption localized (sim + measure) #133 — Spatial × FEC ✅ sim (PR Spatial diversity keeps corruption localized — the SBI precondition (#133) #147): diversity cuts SNR variance → moves frames out of the frame-wide-loss bin into the SBI-salvageable one (a mobility effect; improves SBI's precondition)
• Spatial × frequency — space-frequency interleave for the outer FEC #134 — Spatial × frequency ✅ sim done (PR Space-frequency diversity sim for the outer FEC (#134) #143): frequency = dependable axis, spatial = motion bonus, they multiply
• Spatial × energy — active-chain count as an adaptive-link lever #135 — Spatial × energy ✅ sim (PR Active-chain-count energy lever sim (#135) #145): energy-optimal chain count adapts to fade state — 1 static, up to 4 at a deep mobile link
• Per-temporal-layer spatial UEP (diversity order as a per-layer knob) #136 — Per-temporal-layer spatial UEP ✅ sim (PR Per-SVC-layer spatial UEP: the third protection axis (#136) #146): the 3rd axis already lived in svc_tx.h (STBC on critical/base); now quantified — spatial knob adds ~2.5 dB survival on top of MCS+FEC

Speculative

• TX beam-steer feasibility study for a moving broadcast source (speculative) #137 — TX beam-steer feasibility ✅ (PR Beam-steer feasibility study for a moving broadcast source (#137) #148): NO-GO for a moving source — pointing budget (28° error) puts the beam in a null; feedback-free STBC/MRC wins. Only slow/known-geometry clears STBC

Key finding from #130 — re-prioritisation (2026-07-02)

The on-hardware RX-MRC measurement changed the picture: static spatial-diversity
gain is small (MRC over the best single chain was 2–10% across both 8814 SKUs,
1M/6M, several positions), because the chains' instantaneous channels are
effectively correlated at a fixed position — even for the adapter whose rotation
N_eff was 3.8. Rotation-decorrelation does not translate to a static-link
benefit. The value of spatial diversity is in mobility / fading, where the
best chain constantly changes, plus antenna-selection insurance (~60× vs the worst
chain). This reshapes the roadmap:

1. Linchpin — mobile/fading combining measurement ✅ DONE, thesis CONFIRMED
   (PR Mobile combining measurement: RX-path toggle + mobility analyzer (spatial-diversity #130) #142). Toggling a fixed chain vs all-4 while moving the RX: under motion
   the best chain constantly shifts, so a fixed antenna hits deep-fade dropouts
   that MRC eliminates (half the delivery variance) — static +2 % → moving +17 %
   on the 4-internal 8814. That is the moving-drone FPV case, so the spatial
   techniques (STBC, spatial×FEC, energy lever) are worth pursuing for the
   mobile scenario. Bonus finding: physical antenna spacing decides static
   benefit — wide external dipoles decorrelate even at rest (+17 % static), a
   packed internal array only under motion; this can invert the rotation N_eff
   ranking (spacing > count for real fading).
2. Frequency diversity is immune to the static-correlation problem — different
   channels fade independently regardless of motion. So Spatial × frequency — space-frequency interleave for the outer FEC #134 (space×frequency)
   rises in priority: lead with the frequency axis (hopping, already shipped),
   which delivers real diversity even static.
3. ANTSEL per-frame antenna-selection diversity probe (rescoped from CSD) #131 (antenna selection) is validated as "nearly as good as MRC" at static
   — selection of the best chain captured most of the benefit (MRC added only a
   few %). Simpler than combining; worth the ANTSEL probe.
4. Spatial × energy — active-chain count as an adaptive-link lever #135 (energy lever) gets a sharper thesis — since extra chains add little
   at static, a ground station should run fewer chains to save baseline power
   and light up chains only under fading/motion. Chain count adapts to fade
   state, not just range.

Sequencing (updated)

#128, #129, #130 ✅ → mobile combining measurement ✅ (thesis confirmed,
PR #142) → #134 (frequency-led diversity) + #131 (selection) → #132 (STBC,
now justified for the mobile case) → #133, #135, #136 (composition) → #137.

Current-state anchors

• CSD is chip-validated; STBC lives in the radiotap path but on-air STBC is
  unresolved; 4-path RF is supported for the 8814.