fuse_proj: Initial operational package for 3 researchers (Pavlenko/Blizno/Moroz)

Multimodal fusion research on StripNet+GTA-UAV proxy:
- 3 independent fusion tracks: condition-aware (A), token/bottleneck (B), role-aware (C)
- Shared interfaces, protocol, dataset audit, baseline benchmarks
- Canonical version-chain references to vault (SPEC, ANALYSIS, TRIAGE)
- Personalized task plans and decision tables for each researcher
- 3 generated DOCX task assignment files with milestones and DoD checklist
- Full modality dropout diagnostics and missing-modality robustness requirements
- Data contract, benchmark registry, experiment tracking infrastructure

Operational documents:
- docs/00_project/: MERIDIAN context, protocol, repository reuse guide, experiment specification
- docs/01_tasks/: Master assignment + 3 individual researcher tracks + joint integration
- docs/02_references/: Core literature, version-chain bases, code maps
- docs/03_codebase_guides/: Existing code snapshots from vault
- scripts/: gen_task_plans.js (DOCX generation), placeholder infrastructure
- vendor_reference/: Snapshots of caption_test, depth_edges_annotate, existing SOFIA/SegModel code
- reports/, results/, experiments/: Shared output structure for all 3 researchers

3 DOCX files generated from gen_task_plans.js (Times New Roman 14pt, GOST format):
- План_заданий_Павленко_БВ.docx (Condition-Aware track, fusion API owner)
- План_заданий_Близно_МВ.docx (Token/Bottleneck track, benchmark owner)
- План_заданий_Мороз_ЕС.docx (Role-Aware track, data contract owner)

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>

docs/02_references/03_segmentation/_version_chain/DELTA_pair_A_seg_revised.md

@@ -0,0 +1,705 @@
+---
+type: delta
+status: draft
+date: 2026-05-06
+parent: "[[REVIEW_segmentation_pairA]]"
+related:
+  - "[[../00_overall/SPEC_fusion_ACF_MERIDIAN]]"
+  - "[[MASTER_synthesis_cached_tensors]]"
+  - "[[DELTA_E1_pair_D_text_fusion]]"
+  - "[[DELTA_E1_text_encoder_F_variants]]"
+  - "[[HYP_MambaVision_OV_segmentation]]"
+tags:
+  - delta
+  - decision/delta
+  - component/cvgl
+  - method/film
+  - method/segmentation
+  - arch/dinov3
+  - gate/E1
+  - priority/high
+phase: E1
+hypotheses_added:
+  - H_pair_A_5
+  - H_pair_A_6
+  - H_pair_A_7
+  - H_pair_A_8
+author: claude
+---
+
+# DELTA §2.6 — Pair A semantic fusion (cached uint8 → categorical embedding)
+
+> [!summary] TL;DR
+> **Major architectural shift**: anchor REVIEW_segmentation_pairA (FC-CLIP-style **frozen-provider + Mona-LoRA + mask-pooled descriptor** через SegEarth-OV3 backbone features) **superseded** новым cached tensors anchor ([[MASTER_synthesis_cached_tensors]] §3): **cached uint8 segm tensor [1,256,256] ∈ [0,16] → nn.Embedding(17,64) с LRSCLIP-PCA-64 warm-start → 4-layer conv encoder (1→32→64→96→128) → GGeM → FiLM-A 128→256→(2×1024×5), ~2.3M trainable**. Mona-LoRA-A **deprecated** — нет SegEarth-OV3 backbone features в pipeline (только dense class IDs).
+>
+> 5 refinements (D1-D5): D1 confirm categorical-embed > mask-pooled (forced by cached tensors regime — нет raw features); D2 embed init ablation **LRSCLIP-PCA-64** vs sentence-template averaging vs frozen-LRSCLIP vs random; D3 cardinality 17 anchor (SegEarth-OV3 native), evaluate iSAID-21 / LoveDA-7 trade-off; D4 conv encoder depth 4-layer оптимально (verify against 3 / 5 / 6); D5 §0.8 categorical-embedding-aware FiLM-альтернативы — **per-class FiLM (E.11 MoFiLM) добавляется как F3-research-A** (γ,β routed по indicator класса); aux seg head на student — risk polluting features → **reject as primary, defer to E5 ablation**.
+>
+> **4 новые гипотезы** H_pair_A_5..8: warm-start ablation, per-class FiLM gain, cardinality robustness, aux seg loss feature pollution risk. **H_fus_A_1..4 status updates**: H_fus_A_1 структурно изменился (Mona-LoRA → nn.Embedding), H_fus_A_2 still relevant (DINOv3 ViT-L/16 anchor), H_fus_A_3 superseded (co-training out of scope в cached tensors), H_fus_A_4 invalid (binary-stack vs argmax — cached anchor argmax-class-IDs).
+
+---
+
+## §1. AS-IS — anchor состояние (revised cached tensors regime)
+
+### 1.1. Anchor architecture (MASTER_synthesis_cached_tensors §3, 2026-04-20)
+
+> [!info]+ Pair A revised — cached uint8 → categorical embedding pipeline
+
+**Provider:** SegEarth-OV3 (SAM 3 + PE-L+, 17-class taxonomy, training-free OVSS, mIoU 53.4% avg на 8 RS benchmarks). Pre-computed offline → cached uint8 tensors [1,256,256] ∈ [0,16] в SafeTensors на NVMe SSD (130 KB/pair).
+
+**Архитектура fusion (~2.3M trainable):**
+
+```
+seg [1,256,256] uint8 ∈ [0,16]
+    │
+    ▼
+nn.Embedding(17, 64)              ← warm-start через LRSCLIP-PCA-64
+    │
+    ▼
+[64, 256, 256] feature map
+    │
+    ▼
+4-layer conv encoder:
+  conv 7×7 stride 2 (64→64)       — first downsample
+  conv 3×3 stride 2 (64→96)
+  conv 3×3 stride 2 (96→128)
+  conv 3×3 stride 2 (128→128)
+    │
+    ▼
+[128, 16, 16] feature map
+    │
+    ▼
+GGeM pool (learnable p-exponent)
+    │
+    ▼
+[128] descriptor
+    │
+    ▼
+FiLM-A head MLP 128→256→(2×1024×5)
+    │
+    ▼
+γ_A^{(20-24)}, β_A^{(20-24)}     → blocks 20-24 DINOv3 ViT-L/16
+                                  через shared 256-d bottleneck
+```
+
+**Параметрический бюджет** (verified MASTER §3):
+- nn.Embedding(17, 64) = 1088 params (≈ 0)
+- conv encoder = ~370K
+- FiLM-A MLP = ~2M
+- **Total Pair A trainable = ~2.3M**
+
+**Warm-start procedure:**
+1. 17 prompts из `seg_classes.py UNIFIED_PROMPTS` (e.g., "a satellite image of urban area", "a satellite image of dense forest", "a satellite image of water body", ...)
+2. Encode через LRSCLIP ViT-L/14 (768-dim output)
+3. PCA-decompose → 64 components → [17, 64] init weights
+4. nn.Embedding remains learnable во время training
+
+**Mona-LoRA-A deprecated** ([[MASTER_synthesis_cached_tensors]] §3 deprecation rationale): в cached tensors режиме нет SegEarth-OV3 backbone features (только class IDs); nn.Embedding с warm-start уже produces learnable semantic space, дополнительный adapter избыточен.
+
+### 1.2. Anchor justification (компактно)
+
+- **Cached tensors regime** ([[MASTER_synthesis_cached_tensors]] §1-2): provider features заменены на pre-computed tensors → **structural enforcement** на mask-pooled fusion (mask-pool over **what features**? нет raw seg features в pipeline)
+- **17-class taxonomy** = SegEarth-OV3 native (background, urban, forest, water, road, building, field, bare earth, wetland, etc.) — natural fit для Polyakova caption inventory whitelist (DELTA §2.4 D1 terrain inference)
+- **GGeM pool** = anchor для feature aggregation (cross from F11, [[../../1_lit_research/_reviews/СИНТЕЗ_3_fusion]] §6)
+- **FiLM-A → blocks 20-24** = same injection point как Pair B/C/D/E (5-way unified)
+- **No additional Mona-LoRA**: parameter budget стал ~2.3M вместо anchor ~5-10M (REVIEW_segmentation_pairA §9 H_fus_A_1)
+
+### 1.3. H_fus_A_1..4 (anchor) — updated status
+
+| ID | Original (REVIEW_segmentation_pairA) | New status (DELTA 2026-05-06) | Rationale |
+|:--|:--|:-:|:--|
+| H_fus_A_1 | Frozen-provider + Mona-LoRA + mask-pooled → ΔR@1 ≥ 2.5% | **Structurally superseded** — replaced by H_pair_A_5..8 | Cached tensors → mask-pool неприменим (нет raw seg features) |
+| H_fus_A_2 | DINOv3 ViT-L/16 vs DINOv2/14+Perceiver, R@1 не упадёт > 0.5% | **Confirmed (still relevant)** | DINOv3 ViT-L/16 = anchor backbone, patch-size match |
+| H_fus_A_3 | Co-training PCGrad LoRA both → ΔR@1 ≤ +0.5% but distillability -0.10 | **Superseded** — out of scope в cached tensors | SegEarth-OV3 frozen offline, нет co-training opportunity |
+| H_fus_A_4 | Binary K=17 mask-stack > argmax + embedding-lookup, ΔR@1 ≥ 0.7% | **Invalid (anchor inversion)** | Cached anchor uses **argmax class IDs** (uint8 [0,16]); binary stack would require K-channel float storage = ×17 storage cost |
+
+---
+
+## §2. Лит-обзор: новые свидетельства (2025-2026)
+
+### 2.1. SegEarth-OV3 (arXiv:2512.08730, Dec 2025) — provider details verified
+
+> [!cite] Источник
+> [arXiv:2512.08730](https://arxiv.org/abs/2512.08730) · GitHub earth-insights/SegEarth-OV-3 · Apache-2.0
+
+**Что нового (verified):**
+- SAM 3 + PE-L+ backbone (Bolya et al. arXiv:2504.13181), input 1008×1008 stride-16
+- **Training-free** inference через dual-head mask fusion (semantic + instance) + presence-guided filtering
+- mIoU avg на 8 RS benchmarks: **53.4%** (vs 40.7% CorrCLIP, 39.1% CAT-Seg)
+- Per-dataset ranges: WHU-Aerial 86.9%, Cityscapes 69.7%, LoveDA 47.4%, iSAID (~50%)
+
+**Implication для DELTA:**
+- 17-class anchor taxonomy = **SegEarth-OV3 native protocol** — confirmed
+- Provider stride-16 = native match с DINOv3 ViT-L/16 (patch 16) — no resampling needed
+- **Open-vocabulary prompt drift** (REVIEW §10) — risk при обновлениях provider; mitigated через cached tensors **frozen после version-pinning**
+
+### 2.2. FarSLIP v2 (arXiv:2511.14901, Nov 2025) — Pair A redirect (DELTA §2.5 D7)
+
+> [!cite] Источник
+> [arXiv:2511.14901](https://arxiv.org/abs/2511.14901) · GitHub NJU-LHRS/FarSLIP
+
+**Что нового:**
+- **Patch-to-patch distillation** (vs стандарт patch-to-CLS) — улучшает feature discriminability при сохранении semantic coherence
+- **MGRS-200k dataset** — first multi-granularity RS dataset с object-level text supervision
+- **SOTA на 8 RS OVSS benchmarks** + zero-shot classification + image retrieval
+- Только ViT-B/16 (512-dim text encoder)
+
+**Implication для DELTA:**
+- **FarSLIP v2 — alternative provider candidate** для Pair A (vs SegEarth-OV3 anchor)
+- **Decision: НЕ заменять SegEarth-OV3** — DELTA §2.5 D7 redirect уже отметил: FarSLIP полезен как **interpretation alternative**, но его 200K dataset значительно меньше SegEarth-OV3 (built on SAM 3, training-free на all open-vocab queries)
+- **Backlog research direction**: FarSLIP-augmented warm-start для nn.Embedding(17,64) через MGRS-200k object-level text → может улучшить semantic discriminability vs LRSCLIP scene-level prompts
+
+### 2.3. CLIP-based semantic segmentation 2025 trends (arXiv:2503.20826, Mar 2025)
+
+> [!cite] Источник
+> «Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation» · Mar 2025
+
+**Что нового:**
+- DenseCLIP / MaskCLIP / ComCD trends — CLIP dense knowledge mining для seg-сигнала
+- **CoOp / CLIP-Adapter** — lightweight trainable prompts/adapters insert на frozen CLIP
+- **Class Token Fusion (ICTF) + Patch Token Fusion (PTF)** modules — dual-branched converter
+
+**Implication для DELTA:**
+- ICTF/PTF — alternative для FiLM-conditioning (категории K.1 modality token concat в §0.8 chklist)
+- **Decision: НЕ принимать как primary** — наш FiLM-A анкор более параметр-эффективен (~2.3M vs ICTF/PTF ~10-30M)
+- Add to research backlog для post-E1 ablation (F3-research-A alternative)
+
+### 2.4. Vision-Language Weed Segmentation (arXiv:2602.23677, 2026)
+
+> [!cite] Источник
+> «Vision-Language Semantic Grounding for Multi-Domain Crop-Weed Segmentation»
+
+**Что нового:**
+- **FiLM layers conditioned on natural language captions** для UAV agricultural seg
+- Channel-wise feature refinement при сохранении fine-grained spatial localization
+- **CAGE module** = dual-path cross-attention gated enhancement + global FiLM modulation + spatial context refinement
+
+**Implication для DELTA:**
+- **Direct precedent для FiLM-conditioning по text для UAV seg** — confirms anchor pattern (FiLM-A modulates DINOv3 blocks 20-24 conditioned on segmentation embedding)
+- Anchor ✅ confirmed; CAGE module — secondary alternative для Architecture B (Hybrid)
+
+### 2.5. SkySense V2 (arXiv:2507.13812, ICCV 2025) — RS foundation alternative
+
+> [!cite] Источник
+> [arXiv:2507.13812](https://arxiv.org/abs/2507.13812) · ICCV 2025
+
+**Что нового:**
+- Unified ViT + MoE + APM (Adaptive Prompt Modulation) для RS foundation
+- +1.8 pp avg vs SkySense (NMI 2025)
+- Co-training paradigm (SegEarth-OV3 + DINOv3 + MoE)
+
+**Implication для DELTA:**
+- Reject as Pair A provider — SkySense V2 разрушает cached tensors decoupling (co-training requires online seg)
+- **Acquired backlog B7** ([[../../1_lit_research/_reviews/READING_BACKLOG]] §2.5 LOW) — future reference только
+
+### 2.6. Categorical embedding init landscape
+
+| Init strategy | Method | Cost | Expected benefit | Source |
+|:--|:--|:-:|:--|:--|
+| **LRSCLIP-PCA-64** (anchor) | 17 prompts → LRSCLIP encode → PCA → [17, 64] | 1× LRSCLIP forward + PCA (~30 sec) | semantic warm-start, +1-2pp R@1 vs random | MASTER §3 (anchor) |
+| Sentence-template averaging | n templates per class avg → encode → PCA | n× LRSCLIP forwards | template robustness, marginal +0.3-0.5pp | CoOp / DenseCLIP precedent |
+| Frozen-LRSCLIP (no PCA) | encode → frozen [17, 768] (no learnable Embedding) | 0 train cost on Embedding | no fine-tuning capability на embedding space | reject — anchor needs learnable |
+| DINOv3-text init | DINOv3-text не существует (DINOv3 = vision-only) | N/A | N/A | reject — N/A |
+| FarSLIP v2 init | FarSLIP encode → PCA → [17, 64] | similar cost | object-level discrim better than scene-level | research direction (B6 backlog) |
+| Random init (cold-start) | init.normal_(mean=0, std=0.02) | 0 cost | slow convergence (need full E1 to reach warm-start level) | reject — anchor evidence (REVIEW_text_pairD §10 H_tensor_1: warm-start > cold-start) |
+
+---
+
+## §3. DELTA — что изменяется vs anchor
+
+### 3.1. Что НЕ меняется (anchor сохраняется)
+
+| Anchor | Источник | Действие |
+|:--|:--|:--|
+| Cached uint8 segm tensor [1,256,256] ∈ [0,16] | MASTER §2 | ✅ keep |
+| 17-class taxonomy SegEarth-OV3 native | SegEarth-OV3 paper | ✅ keep |
+| nn.Embedding(17, 64) categorical | MASTER §3 | ✅ keep |
+| LRSCLIP-PCA-64 warm-start | MASTER §3 | ✅ keep (D2 verifies vs alternatives) |
+| 4-layer conv encoder (1→32→64→96→128) | MASTER §3 (note: anchor пишет 1→32... но input уже 64-channel после embedding, верный — 64→64→96→128→128) | ✅ keep with notation fix |
+| GGeM pool (learnable p) | MASTER §3 + F11 anchor | ✅ keep |
+| FiLM-A MLP 128→256→(2×1024×5) | MASTER §3 | ✅ keep |
+| ~2.3M trainable params budget | MASTER §3 | ✅ keep |
+| Blocks 20-24 DINOv3 injection | REVIEW §9 | ✅ keep (5-way unified) |
+| Shared 256-d bottleneck | MASTER §3 | ✅ keep |
+| Mona-LoRA-A deprecated | MASTER §3 | ✅ keep deprecated |
+| Co-training out of scope | MASTER §3 + DELTA §2.3 anchor | ✅ keep frozen |
+
+### 3.2. Что предлагается УТОЧНИТЬ (Decision DELTA Table)
+
+| # | Item | Было (anchor) | Станет (DELTA) | Threshold для acceptance | Источник evidence |
+|:-:|:--|:--|:--|:--|:--|
+| **D1** | Categorical embedding vs feature-pooled fusion | implicit anchor | **Explicit confirm**: cached tensors regime structurally enforces categorical (no raw features available); H_fus_A_1 superseded by H_pair_A_5 | structural decision: 130 KB/pair uint8 storage vs ~30 MB/pair float16 K=17 binary stack (×230 cost) | MASTER §2 storage budget |
+| **D2** | Embedding init = LRSCLIP-PCA-64 | anchor (single strategy) | **3-way ablation**: LRSCLIP-PCA-64 (anchor) vs sentence-template averaging (3 templates per class) vs random cold-start | R@1(anchor warm-start) ≥ R@1(cold-start) + 1.0pp | H_pair_A_5 (NEW) |
+| **D3** | Cardinality 17 (SegEarth-OV3 native) | anchor | **17 primary; iSAID-21 / LoveDA-7 evaluated в E5 ablation** для cross-dataset transfer | R@1(17) ≥ R@1(7) + 0.5pp; R@1(17) ≥ R@1(21) − 0.5pp (parity) | H_pair_A_7 (NEW) |
+| **D4** | Conv encoder depth = 4 layers | anchor | **Confirm 4 optimal vs 3 / 5 / 6 ablation** (preparatory probe) | param budget proportional vs R@1 marginal gain ≤ 0.3pp diff | H_pair_A_8 (subset) |
+| **D5** | Per-class FiLM (γ,β routed by class) — implicit single-stream | global GGeM single FiLM-A | **Add F3-research-A: per-class FiLM (E.11 MoFiLM)** — 17 separate γ,β heads, gated by class indicator | research-only, post-E1 conditional на anchor wins | H_pair_A_6 (NEW) |
+| **D6** | Aux seg loss on student | not in anchor | **Reject as primary, defer to E5 ablation** (risk polluting features) | aux seg loss adds CE/Dice gradient on student → potential negative transfer; defer | H_pair_A_8 (NEW) |
+| **D7** | FarSLIP v2 alternative provider | not considered | **Track as backlog research** (B1 P1 in §2.5 backlog) — defer post-E1 | future ablation для embedding init alternative (FarSLIP-PCA-64 vs LRSCLIP-PCA-64) | DELTA §2.5 D7 redirect + §2.6 (b) |
+
+### 3.3. Что предлагается ДОБАВИТЬ
+
+#### 3.3.1. Per-class FiLM (F3-research-A) — Mixture-of-FiLM (MoFiLM) категория E.11 §0.8
+
+**Rationale:** anchor использует **single FiLM-A** на global GGeM descriptor — теряет per-class granularity. **F3-research-A**:
+
+```
+seg [1,256,256] uint8 ∈ [0,16]
+    │
+    ▼ embedding (17, 64) + warm-start
+[64, 256, 256]
+    │
+    ▼ class indicator vector c ∈ {0,1}^17 (presence per class в image)
+    │
+    ▼ FiLM-A_c heads (17 separate MLPs)
+γ^c, β^c ∈ R^{2×1024×5} per class c
+    │
+    ▼ gated по indicator: γ_A = Σ_c c_c · γ^c / Σ c_c
+    │
+    ▼ same → blocks 20-24
+```
+
+**Cost:** 17× larger MLP head budget = ~17 × 2M = **~34M trainable** (vs anchor 2M); может быть оптимизировано через **shared 64-d bottleneck** (17 × 0.5M = 8.5M).
+
+**Threshold:** R@1 ≥ R@1(anchor) + **2pp** (compensates 4-15× param overhead).
+
+**Decision:** **research-only**, post-E1 conditional на anchor wins; не блокирует E1 launch.
+
+#### 3.3.2. Sentence-template averaging warm-start (D2)
+
+**Procedure:** для каждого из 17 классов:
+- 3 prompts (e.g., "a satellite image of urban area", "an aerial photo of urban district", "a top-down view of urban region")
+- Encode each через LRSCLIP, average → [768]
+- PCA all 17 → [17, 64]
+
+**Expected:** template diversity reduces caption-to-prompt mismatch при inference; +0.3-0.5pp R@1 vs single-prompt anchor.
+
+**Cost:** 3× LRSCLIP forward at warm-start (~90 sec total, one-time).
+
+#### 3.3.3. Cardinality ablation (D3)
+
+**Procedure (E5 ablation, post-E1):**
+- 17 anchor (SegEarth-OV3 native)
+- 21 iSAID extension (15 foreground + 1 background = 16, plus SegEarth-OV3 unmapped 5 = 21)
+- 7 LoveDA reduced (6 foreground + 1 background)
+
+**Expected:**
+- 17 anchor optimal — balanced fine-grained vs taxonomy stability
+- 7 reduced — coarser, may improve robustness to class confusion
+- 21 expanded — finer, may improve discriminability но expand provider re-training cost
+
+### 3.4. Conflicts с anchor
+
+> [!warning]+ Conflict 1 — REVIEW_segmentation_pairA H_fus_A_1..4 vs cached tensors regime
+>
+> REVIEW _segmentation_pairA канон (2026-04-20) — Mona-LoRA-A + mask-pooled descriptor (FC-CLIP-style); MASTER_synthesis_cached_tensors §3 (2026-04-20) — nn.Embedding(17,64) + LRSCLIP-PCA-64. **Two anchors из одной даты противоречат друг другу**. **Resolution:** MASTER_synthesis_cached_tensors **supersedes** REVIEW для cached tensors regime (новая major revision). REVIEW сохраняется как **historical reference** для FC-CLIP-style approach в случае cached tensors regime fails (например, R_provider — provider drift makes pre-computed tensors stale).
+
+> [!warning]+ Conflict 2 — anchor пишет "1→32→64→96→128" но nn.Embedding выход = 64-channel
+>
+> MASTER §3 текст: «conv 7×7 stride 2 c 1→32 channels для начального feature extraction, затем три conv 3×3 stride 2 c 32→64→96→128». Но input уже [64, H, W] после nn.Embedding(17,64). **Notation inconsistency**. **Resolution:** correct архитектура должна быть либо (a) **64→64→96→128→128** (4 layers с initial channel=64), либо (b) **1×1 reduction** 64→1 в начале + (1→32→64→96→128) — но это теряет embedding semantic. **DELTA D8: notation fix** — verify в implementation что используется (a). [Cross-link: MASTER §3 line 73].
+
+> [!info]+ No conflict — Pair A 9-zone L2 caption coordinate system
+>
+> DELTA §2.4 H_caption_2 (9-zone grid) использует SAME coordinate system как Pair A 9-zone seg quadrant maps (Polyakova MapTextualizer, [[PIPELINE_text_annotation_full]] §3.2). Anchor confirmed sync.
+
+### 3.5. Risks of refinement
+
+> [!danger]+ R1 — SegEarth-OV3 provider drift при обновлениях
+>
+> Anchor REVIEW §10 already noted prompt-set drift risk. **Cached tensors mitigates** через offline pre-compute + version-pinning (hash-suffix SafeTensors filenames). But: при обновлении SegEarth-OV3 (next release) **все 962K tensors stale** → re-compute cost ~100 H100-hours. **Mitigation:** strict version-pinning + 500-image regression test set + automated drift detection.
+
+> [!warning]+ R2 — Embedding cold-start collapse
+>
+> If LRSCLIP-PCA-64 warm-start ablation (D2) shows random cold-start ≈ warm-start, anchor warm-start cost (1× LRSCLIP forward + PCA, ~30 sec) **не оправдан**. **Mitigation:** if H_pair_A_5 disproven, simplify pipeline (drop warm-start, use random init).
+
+> [!warning]+ R3 — Per-class FiLM (F3-research-A) param explosion
+>
+> 17 separate FiLM heads × 2M = 34M trainable (vs 2M anchor) — **15× overhead**. Memory 4090 borderline. **Mitigation:** shared bottleneck factorization (17 × 0.5M = 8.5M) — reduces to 4× overhead; conditional на E1 anchor wins, defer.
+
+> [!danger]+ R4 — Aux seg head на student polluting features
+>
+> Если в Student добавить auxiliary seg head (CE loss на cached uint8 → class-IDs), gradient seg-task может перекрывать CVGL-task gradient (REVIEW §8.3 negative transfer evidence). **Mitigation D6:** **reject aux seg head** в Student-side; if needed, evaluate **только в E5 ablation** с PCGrad gradient surgery (REVIEW §8.3).
+
+> [!warning]+ R5 — 1×1 notation conflict (Conflict 2)
+>
+> Если implementation использует (1→32→64→96→128) вместо (64→64→96→128→128), теряется embedding-injected semantic content (1×1 reduction до single channel). **Mitigation D8:** code review fix BEFORE E1 launch — verify pipeline_256.py / pair_A_module.py implementation.
+
+### 3.6. Отвергнутые предложения
+
+> [!failure]+ Reject — Mona-LoRA-A revival
+>
+> Anchor MASTER §3 deprecated Mona-LoRA-A для cached tensors. Adding back требует **reintroducing online SegEarth-OV3 provider** для feature extraction → defeats cached tensors purpose (~100h H100 inference вместо 0). Reject.
+
+> [!failure]+ Reject — Binary K=17 mask stack (vs argmax uint8)
+>
+> H_fus_A_4 (REVIEW §9) hypothesizes binary stack > argmax. **Counter-evidence**: storage 17× cost (130 KB → 2.2 MB per pair = ~2 TB extra на World-UAV); compute overhead для conv encoder K-channel input (×17 first layer). **Threshold для acceptance в anchor**: ΔR@1 ≥ 5pp justify storage cost. Anchor evidence: 0pp expected (uint8 + nn.Embedding lookup retrieves identical semantic vector). Reject.
+
+> [!failure]+ Reject — SkySense V2 / GeoLangBind alternative provider
+>
+> Both require **co-training** (SkySense V2 unified ViT + MoE + APM; GeoLangBind agglomerative VL) → break cached tensors decoupling. Out of scope в cached tensors regime. Reject.
+
+> [!failure]+ Reject — Co-training PCGrad LoRA both DINOv3 + PE-L+ (H_fus_A_3)
+>
+> H_fus_A_3 (REVIEW §9) was conditional on FC-CLIP-style anchor. New cached tensors anchor — provider **frozen offline**, no co-training opportunity. Even if extracted, PCGrad gradient surgery overhead (×1.5 compute) → not justified. Reject.
+
+> [!failure]+ Reject — Aux seg head на student (D6)
+>
+> Adding CE/Dice loss на student adds gradient interference (REVIEW §8.3). **Empirical evidence** Polyakova v8 production iteration — adding altitude-reasoning to prompt каскадом **+73% FDR** (DELTA §2.4 §3.6 reject 4×4 grid analogous). Multi-task learning без PCGrad → expected feature pollution. Defer to E5 ablation only.
+
+> [!failure]+ Reject — Random cold-start (no warm-start)
+>
+> H_tensor_1 (MASTER §6 hypothesis): warm-start nn.Embedding via LRSCLIP > cold-start random (Medium confidence, testable in E3 ablation). Anchor preserves warm-start. Reject random as primary; keep as ablation comparison.
+
+---
+
+## §4. §0.8 FiLM-АЛЬТЕРНАТИВЫ — обязательный чек-лист (categorical-embedding-aware)
+
+> [!important] §0.8 mandatory для fusion-prompt §2.6
+>
+> Categorical-embedding-aware variants особенно: B.5 Q-Former, F.1 MoE per-class, K.1 modality token concat.
+
+### 4.1. Категории A-L: status в DELTA
+
+| Cat | Метод | Anchor decision | Reason |
+|:-:|:--|:-:|:--|
+| **A.1** Early concat (input-level) | Reject | RGB+seg input concat нарушает DINOv3 patch-embed weights |
+| **A.2** Late concat | Reject | Linear fusion, weak |
+| **B.5** Q-Former (BLIP-2) | Architecture C research-only | Bottleneck destroys dense correspondence (как в Pair D §2.3 reject as primary) |
+| **C.1** GMU (Gated Multimodal Unit) | Reject | Gating only, no shift component (FiLM superset) |
+| **E.7** Conditional FiLM (anchor) | **PRIMARY** | Anchor ✅ |
+| **E.11** Mixture-of-FiLM (MoFiLM) | **F3-research-A candidate** (D5) | Per-class FiLM, 17 separate γ,β heads gated by indicator |
+| **F.1** Top-k MoE per modality | Reject as primary | Modality-level MoE — для §2.9, не Pair A class-level |
+| **F.3** Soft MoE (Puigcerver 2024) | Research direction | Per-class soft routing — alternative для F3-research-A |
+| **G.1** Houlsby adapter | Reject | Mona deprecated; Houlsby не дает benefit над FiLM в cached |
+| **G.3** LoRA | Already in use | LoRA на DINOv3 blocks 20-24 — anchor |
+| **K.1** Modality token concat (BEiT-like) | Reject | Adding `[SEG]` token к sequence требует full retrain DINOv3 |
+| **L.3** MultiLevelFiLM + MoE routing | F3-research-A alternative | Same as E.11 effectively |
+
+### 4.2. DELTA-чеклист для F3-research-A (per-class MoFiLM)
+
+- [x] **Категория A-L?** E.11 Mixture-of-FiLM
+- [x] **Params overhead vs anchor (~2.3M)**: ~34M (15×) или ~8.5M с shared bottleneck (4×)
+- [x] **Identity-at-init свойство сохраняется?** Да — tanh-α init=0 + per-class γ^c initialized normal
+- [x] **Что происходит при class-absent inference?** Indicator c_c=0 → corresponding γ^c, β^c исключаются из weighted sum; graceful degradation
+- [x] **INT8-friendly?** Per-class routing add discrete branching → INT8 calibration challenges; defer to E9 quantization phase
+- [x] **Совместимо ли с frozen DINOv3-L?** Да — applied in shared 256-d bottleneck downstream
+- [x] **К-5 scalability**: independent of K=5 modality count; uses K=17 class count internally
+- [x] **Источник идеи**: §0.8 E.11 + per-class adapter precedent (CoOp / DenseCLIP)
+- [x] **Ablation план**: F3-research-A только post-E1 conditional на anchor wins; E5 ablation track
+
+---
+
+## §5. CVGL DOMAIN AWARENESS — чек-вопросы (MEDIUM domain-aware §0.6)
+
+§2.6 — **MEDIUM domain-aware** ([[ПРОМПТЫ_анализ_литобзоров_по_задачам_MERIDIAN]] §1.1). Pair A fusion должен handle классы partially-overlapping для sat/UAV.
+
+| Категория | Status |
+|:--|:--|
+| **F1 scale** | ✅ partially — class-IDs не зависят от scale; 17-class taxonomy включает scale-invariant categories (urban / forest / water); SegEarth-OV3 stride-16 native match с DINOv3 |
+| **F2 viewpoint mismatch** | ⚠ partially — sat nadir / UAV oblique классы differently visible (footprint vs facade); SegEarth-OV3 inferred per-pixel handles per-view; aug consistency through cached tensors |
+| **F4 illumination / sun angle** | ✅ classes invariant к sun angle; cached tensors frozen — no augmentation required |
+| **F5 repetitive texture** | ✅ class-IDs collapse repetitive regions (urban / building grid → "building"); per-zone fusion (Pair A 9-zone aggregation) restores spatial discriminability |
+| **F6 occlusion / dynamic objects** | ⚠ - dynamic objects (cars, pedestrians) NOT in 17-class SegEarth-OV3 taxonomy → naturally excluded; static objects (buildings) preserved |
+| **#13 GPS noise tolerance** | N/A — Pair A class-IDs invariant к GPS noise |
+| **#14 temporal mismatch** | ✅ taxonomy-level seg robust к seasonal changes (forest стабильно vs sun-leaves) |
+
+**Conclusion:** MEDIUM domain-aware addressed. Pair A naturally handles 5/7 categories ✅, 2/7 ⚠ (viewpoint mismatch + occlusion) handled через provider stride-16 + 17-class filtering.
+
+---
+
+## §6. GAP ANALYSIS — backlog для §2.6
+
+### Output: Таблица A — В vault, требуют углубления (SHALLOW)
+
+| # | Paper / Author Year | Paper-ID | Существующая заметка | Глубина | Priority | Est. MODE-A time |
+|:-:|:--|:--|:--|:--|:-:|:-:|
+| A1 | F5 SegEarth-OV3 | F5 | [[CVGL/1_lit_research/2_foundation_models/_legacy/F5_BB_2025_SegEarth-OV3 Exploring SAM 3 for Open-Vocabulary Semantic Segmentation in Remote Sensing Images]] | DEEP | — | — |
+| A2 | F11 GGeM | F11 | [[F11_2022_Group Generalized Mean Pooling for Vision Transformer]] | проверить depth | P2 | 1h |
+| A3 | FarSLIP v2 update | F5 (3_fusion) | [[F5_2025_FarSLIP Discovering Effective CLIP Adaptation for Fine-Grained Remote Sensing Understanding]] | SHALLOW (pre-v2 stub) | P1 (already in §2.5 backlog #29) | — |
+| A4 | F17 SegMAN | F17 (3_fusion) | [[F17_2025_SegMAN Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation]] | SHALLOW | P2 (already in §2.1 backlog #9) | — |
+| A5 | F18 SegDINO | F18 (3_fusion) | [[F18_2025_SegDINO An Efficient Design for Medical and Natural Image Segmentation with DINO-V3]] | проверить depth | P2 | 1h |
+| A6 | LMVMamba | F16 (3_fusion) | [[F16_2025_LMVMamba A Hybrid U-Shape Mamba for Remote Sensing Segmentation with Adaptation Fine-Tuning]] | SHALLOW (already in §2.1 backlog #17) | P1 (HIGH для §2.6) | 1.5h |
+
+### Output: Таблица B — НЕ в vault, требуют acquisition + MODE-A
+
+| # | Paper / Author Year | Title | DOI / arXiv | Status | Priority | Acquisition path |
+|:-:|:--|:--|:--|:--|:-:|:--|
+| B1 | **CoOp / CLIP-Adapter** | Context Optimization for Vision-Language Models | [arXiv:2109.01134](https://arxiv.org/abs/2109.01134) (CoOp) + arXiv:2110.04544 (CLIP-Adapter) | NOT_FOUND | P2 | acquire — embedding init reference |
+| B2 | **DenseCLIP** | Language-Guided Dense Prediction with Context-Aware Prompting | [arXiv:2112.01518](https://arxiv.org/abs/2112.01518) | NOT_FOUND | P3 | acquire — context-aware prompting |
+| B3 | **MaskCLIP / MaskCLIP++** | Open-Vocabulary Universal Image Segmentation with MaskCLIP | [arXiv:2208.12262](https://arxiv.org/abs/2208.12262) (MaskCLIP++) | mentioned in REVIEW §1 | NOT_FOUND deep-dive | P3 | acquire summary |
+| B4 | **VL-Weed Segmentation (2026)** | Vision-Language Semantic Grounding for Multi-Domain Crop-Weed | [arXiv:2602.23677](https://arxiv.org/abs/2602.23677) | NOT_FOUND | P3 | acquire — FiLM-conditioned UAV seg precedent |
+| B5 | **Stepping Out of Similar Semantic Space (ICCV 2025)** | Open-Vocabulary Segmentation | openaccess.thecvf.com/content/ICCV2025/papers/Liu_Stepping_Out... | NOT_FOUND | P3 | acquire — OVSS reference |
+| B6 | **SkySense V2 (ICCV 2025)** | Unified Foundation Model for Multi-modal RS | [arXiv:2507.13812](https://arxiv.org/abs/2507.13812) | NOT_FOUND (already in §2.5 backlog #35) | P3 | low priority — out of cached scope |
+| B7 | **AnyUp (Oct 2025)** | feature-agnostic upsample | [arXiv:2510.12764](https://arxiv.org/abs/2510.12764) | mentioned in REVIEW §7.5 | NOT_FOUND | P3 | acquire — pixel-level fusion alternative |
+
+### Output: Сводная статистика
+
+- Всего цитируемых работ по теме промпта (Pair A seg fusion): ~25
+- DEEP в vault: 1 (F5 SegEarth-OV3) + REVIEW_segmentation_pairA + MASTER §3
+- SHALLOW: 5 (F11, F17, F18, F16/LMVMamba, F5/FarSLIP partial)
+- NOT_FOUND: 7 (CoOp, DenseCLIP, MaskCLIP++, VL-Weed, Stepping Out, AnyUp)
+- **P0 backlog: 0** — anchor MASTER §3 + REVIEW обеспечивает coverage
+- **P1 backlog: 1** (LMVMamba HIGH для §2.6 как cross from §2.1; FarSLIP v2 already in §2.5 backlog)
+- **P2 backlog: 4** (F11 verify, F18 verify, CoOp acquire)
+- **P3 backlog: 5** (DenseCLIP, MaskCLIP++, VL-Weed, Stepping Out, AnyUp)
+
+### Output: Action items
+
+- [ ] **P1:** deepen LMVMamba (already in backlog #17 — promote priority HIGH для §2.6) — H_pair_A_8 evidence
+- [ ] **P2:** verify F11 GGeM depth ([[F11_2022_Group Generalized Mean Pooling for Vision Transformer]]) — anchor reference for pooling
+- [ ] **P2:** verify F18 SegDINO depth — DINO-V3 segmentation precedent
+- [ ] **P2:** acquire CoOp + CLIP-Adapter (B1) — context optimization для embedding warm-start D2 baseline
+- [ ] **P3:** acquire DenseCLIP, MaskCLIP++, VL-Weed Segmentation, Stepping Out (ICCV'25), AnyUp
+
+### Связь с DELTA acceptance
+
+P0 = 0 → **DELTA принимается с full confidence**.
+P1 = 1 (LMVMamba promote) — recommended но не блокирующее.
+P2 = 4 — landscape reading.
+
+---
+
+## §7. Synchronization
+
+### С §2.3 Pair D (text fusion)
+
+- D1 categorical embedding ↔ DELTA §2.3 D1 (F1 hierarchical routing) — **L2 9-zone routing requires Pair A 9-zone seg quadrant grid** as input format (Polyakova MapTextualizer)
+- D5 per-class FiLM → §2.3 §0.8 chklist E.11 MoFiLM secondary (sync confirmed)
+- Anchor blocks 20-24 + 5-way orthogonality → 4-way constraint at Phase 3 включает Pair A (§2.3 anchor curriculum)
+
+### С §2.4 VLM prompts
+
+- D1 terrain inference из seg ↔ DELTA §2.4 D1 (terrain="Unknown" 100% production fix через building >30%, vegetation >60%, water >20%)
+- 17-class taxonomy → Polyakova Anti-Hallucination Object Inventory whitelist (DELTA §2.4 H_caption_3)
+
+### С §2.5 text encoder (F1 anchor refresh)
+
+- LRSCLIP-PCA-64 warm-start uses **LRSCLIP v1** (anchor) — **not DGTRS-CLIP v2**! Verify cosine drift v1↔v2 (DELTA §2.5 H_text_9) — if drift > 0.3, recompute warm-start через DGTRS-CLIP v2
+- **Action D2-extension**: add v2 warm-start as 4th ablation arm if H_text_9 confirms compatibility
+
+### С §2.6 Pair A → §2.7 Pair B (depth) → §2.8 Pair C (CHM) → §2.9 (full 5-modal)
+
+- Unified 4-layer conv encoder design (Pair A 64-channel input differs from Pair B/C/E 1-channel input) — **harmonized** через identical downstream FiLM-X structure
+- 5-way orthogonality regularizer applies к γ_A — Phase 3 unfreezing
+- Pair A semantic + Pair D text — semantic redundancy concern (REVIEW §10): "20 buildings" text vs SegEarth-OV3 building pixel mask. **Mitigation**: orthogonality regularizer suppresses redundant directions (γ_A ⊥ γ_D)
+
+### С §2.10 (KD strategy)
+
+- **Distillability score** для Pair A — anchor мы оцениваем как ✅✅ (FiLM-A signal легко reproducible через scalar context = γ_A=0, β_A=0 при text-drop)
+- F3-research-A per-class FiLM — distillability deg (17 separate heads — student should reproduce all)
+
+### С §2.13 augmentation
+
+- Cached tensors **frozen** — no augmentation на seg uint8 (sync с §2.13 D-rule: только RGB augmented)
+- Geometric aug (rotation, scale) применяется одинаково к RGB + seg cached (consistency requirement)
+
+### С Pavlenko (cross-project)
+
+- Pavlenko's K=3 HID-Fusion — text-only; Pair A semantic outside Pavlenko scope
+- **No cross-project sync needed** для §2.6 (semantic out of Pavlenko scope per [[../../../Павленко_диссертация/2_hypotheses/novelty_differentiation]])
+
+### С Устенко (Object Inventory)
+
+- Устенко DINOv3+LoRA на SynDrone — **alternative provider** для rare-class enrichment (post-E1)
+- 17-class taxonomy SegEarth-OV3 + Устенко rare-class detections → 21-25 class extension в E5 ablation (см. D3 cardinality)
+
+---
+
+## §8. Связь с ROADMAP
+
+### Phase E1 (Teacher 5-modal benchmark)
+
+- Pair A activated в **Phase 3** (epochs 30-40, [[MASTER_synthesis_cached_tensors]] §4)
+- Warm-start LRSCLIP-PCA-64 — preparatory step (~30 sec, ~0 GPU-h)
+- D2 ablation (3-way warm-start vs cold-start vs sentence-template avg): preparatory probe, ~3 runs × 1 epoch = 3-4 GPU-h
+- D3 cardinality ablation: post-E1 / E5 phase (~9 runs × 6h = 54 GPU-h)
+- D5 F3-research-A per-class FiLM: post-E1 conditional, E5+
+
+### H_pair_A — обновлённое resume
+
+| ID | Status | Phase | Notes |
+|:--|:-:|:-:|:--|
+| H_fus_A_1 | **Superseded** by H_pair_A_5 | — | Mona-LoRA → nn.Embedding |
+| H_fus_A_2 | Confirmed (still relevant) | E1 | DINOv3 ViT-L/16 anchor |
+| H_fus_A_3 | **Superseded** | — | Co-training out of scope |
+| H_fus_A_4 | **Invalid** | — | Anchor uses argmax, not binary stack |
+| **H_pair_A_5** *(new, DELTA 2026-05-06)* | High | preparatory | LRSCLIP-PCA-64 warm-start > random cold-start ≥ +1.0pp R@1 |
+| **H_pair_A_6** *(new, DELTA 2026-05-06)* | Medium (research) | E5+ | F3-research-A per-class MoFiLM > anchor single FiLM-A ≥ +2pp R@1 |
+| **H_pair_A_7** *(new, DELTA 2026-05-06)* | Medium | E5 | 17-class anchor ≥ 7-class +0.5pp; ≈ 21-class ±0.5pp parity |
+| **H_pair_A_8** *(new, DELTA 2026-05-06)* | Medium-Low | E5 | Aux seg head на student = -0.5pp R@1 (feature pollution risk) |
+
+(Полные формулировки H_pair_A_5..8 — см. §9.)
+
+### Зависимости / блокировки
+
+- **Блокирует:** §2.9 (5-way fusion synthesis) — Pair A architecture choice → shared bottleneck input dim
+- **Блокируется:** §2.5 F-encoder choice (DGTRS-CLIP v2 may replace LRSCLIP for warm-start), §2.4 caption pipeline (Polyakova MapTextualizer 9-zone seg quadrant)
+- **Sync с:** §2.13 augmentation D-rule (cached tensors frozen)
+
+---
+
+## §9. Новые гипотезы H_pair_A_5..8
+
+### H_pair_A_5: LRSCLIP-PCA-64 warm-start > random cold-start ≥ +1.0pp R@1
+
+**Если** инициализировать nn.Embedding(17, 64) через LRSCLIP-PCA-64 warm-start (17 prompts → LRSCLIP encode → PCA decompose до 64-d → init weights), сохранив learnable Embedding,
+
+**то** anchor warm-start превосходит random cold-start (init.normal_(mean=0, std=0.02)) на **R@1 ≥ +1.0pp** на University-1652 val ПОСЛЕ Phase 3 training (epochs 30-40),
+
+**потому что** (1) H_tensor_1 ([[MASTER_synthesis_cached_tensors]] §6) утверждает warm-start > cold-start с Medium confidence; (2) warm-start даёт meaningful semantic positions для классов forest/building/water → ranges ≈ LRSCLIP semantic space; (3) cold-start требует full Phase 3 (10 epochs) для convergence к meaningful semantics; (4) **3rd ablation arm** sentence-template averaging (3 prompts per class) может give marginal benefit над single-prompt anchor (+0.3-0.5pp).
+
+- **Уверенность:** High
+- **Область:** Embedding init, Phase 3 convergence speed
+- **Baseline:** Random cold-start (init.normal_(mean=0, std=0.02))
+- **Метрика:** R@1 на University-1652 val ПОСЛЕ Phase 3 (epoch 40); also Phase 3 convergence rate (R@1 epoch 30 → 40)
+- **Threshold для успеха:**
+  - R@1(LRSCLIP warm-start) ≥ R@1(cold-start) + 1.0pp
+  - Phase 3 convergence faster (R@1 reached 95% target by epoch 35 vs cold-start epoch 40)
+  - Sentence-template avg ≈ LRSCLIP single-prompt ±0.3pp (parity)
+- **Опровержение:**
+  - ΔR@1 < 0.5pp → warm-start не оправдан, simplify anchor (drop LRSCLIP warm-start, use random)
+  - ΔR@1 < 0% → embedding init irrelevant; skip warm-start
+- **Зависимости:** [[MASTER_synthesis_cached_tensors]] §3 H_tensor_1
+- **Ресурсы:** 3 ablation runs × 3 seeds = 9 runs × 4-6h = 36-54 GPU-h
+- **Эксперимент:** E1.embed_init_ablation (preparatory, before Phase 3)
+
+---
+
+### H_pair_A_6: Per-class FiLM (F3-research-A MoFiLM) > anchor single FiLM-A ≥ +2pp R@1
+
+**Если** заменить anchor single FiLM-A (global GGeM descriptor → 1 MLP → γ,β) на **F3-research-A**: 17 separate FiLM-A_c MLPs (one per class), gated by class indicator c ∈ {0,1}^17,
+
+$$\gamma_A = \frac{1}{|c|_1} \sum_{k=1}^{17} c_k \cdot \gamma^{(k)}, \quad \beta_A = \frac{1}{|c|_1} \sum_{k=1}^{17} c_k \cdot \beta^{(k)}$$
+
+(с shared 64-d bottleneck, total ~8.5M trainable, vs anchor 2M),
+
+**то** F3-research-A даст **R@1 ≥ R@1(anchor) + 2.0pp** на University-1652, оправдывая 4× param overhead,
+
+**потому что** (1) anchor global GGeM averages **все** spatial classes → теряет per-class discriminative weighting; (2) per-class γ,β allows class-specific feature emphasis (e.g., building-image needs different DINOv3 features than water-image); (3) **§0.8 E.11 MoFiLM** evidence — Mixture-of-FiLM ResNet ablation +1.3pp params; (4) cached tensors regime preserves per-image class indicator (uint8 unique values gives c per image) — natural input.
+
+- **Уверенность:** Medium (research)
+- **Область:** Per-class fusion granularity, F3-research-A
+- **Baseline:** Anchor global GGeM single FiLM-A (~2.3M)
+- **Метрика:** R@1 на University-1652 val; per-class breakdown R@1 (urban / forest / water / building)
+- **Threshold для успеха:**
+  - R@1(F3-research-A) ≥ R@1(anchor) + 2.0pp (justifies 4× param overhead)
+  - Per-class R@1 breakdown shows complementarity (different classes win in different metrics)
+- **Опровержение:**
+  - ΔR@1 < 1pp → param overhead не оправдан, anchor sufficient
+  - ΔR@1 < 0pp → per-class routing damages global descriptor, anchor wins
+- **Зависимости:** anchor Phase 3 baseline trained, conditional на anchor wins
+- **Ресурсы:** ~50 GPU-h H100 (4× compute), post-E1
+- **Эксперимент:** E5_research.per_class_FiLM (post-E1 conditional)
+
+---
+
+### H_pair_A_7: Cardinality 17 (SegEarth-OV3 native) ≥ 7-class LoveDA + 0.5pp; ≈ 21-class iSAID ±0.5pp parity
+
+**Если** ablate cardinality {17 SegEarth-OV3 native, 21 iSAID-extended, 7 LoveDA-reduced} с idential downstream architecture (nn.Embedding(K, 64) + warm-start + 4-layer conv + GGeM + FiLM-A),
+
+**то** на University-1652 val:
+- **R@1(K=17) ≥ R@1(K=7) + 0.5pp** (17-class finer discriminability)
+- **R@1(K=17) ≥ R@1(K=21) − 0.5pp** (parity, 21-class no significant gain)
+
+**потому что** (1) 17 classes balance fine-grained vs taxonomy stability; (2) 7-class LoveDA loses critical RS distinctions (forest vs farm-field collapsed); (3) 21-class iSAID adds 4 instance categories (vehicles/ships) — dynamic objects, exclude per Polyakova Tier-3; (4) anchor ✅ confirmed по cross-dataset transfer evidence (SegEarth-OV3 paper Table 1).
+
+- **Уверенность:** Medium
+- **Область:** Cardinality choice, taxonomy stability
+- **Baseline:** K=17 anchor
+- **Метрика:** R@1 на University-1652; cross-dataset transfer (SUES-200, World-UAV)
+- **Threshold для успеха:**
+  - R@1(K=17) ≥ R@1(K=7) + 0.5pp
+  - R@1(K=17) ≥ R@1(K=21) − 0.5pp (parity OR better)
+- **Опровержение:**
+  - K=21 > K=17 by ≥ 1pp → consider iSAID-extended taxonomy (+ provider re-train cost)
+  - K=7 ≥ K=17 → simpler taxonomy preferred
+- **Зависимости:** anchor trained, conditional на post-E1 ablation budget
+- **Ресурсы:** 3 cardinality × 3 seeds = 9 runs × 6h = 54 GPU-h
+- **Эксперимент:** E5.cardinality_ablation
+
+---
+
+### H_pair_A_8: Aux seg head на student polluting features (-0.5pp R@1 risk)
+
+**Если** в Student добавить auxiliary seg head — CE/Dice loss на cached uint8 → student class-IDs prediction (pixel-level seg recovery),
+
+**то** Student R@1 на University-1652 деградирует на **-0.5pp** относительно anchor (no aux seg loss),
+
+**потому что** (1) seg-task gradient (CE/Dice on dense pixel-level) и CVGL-task gradient (image-level retrieval InfoNCE) имеют **разные directions** — без PCGrad одна перекрывает другую; (2) Student ~5M params not enough capacity для совместного multi-task training; (3) aux seg loss adds **feature pollution** — Student forced to retain pixel-class information вместо CVGL discriminative features; (4) precedent — REVIEW_segmentation_pairA §10 "negative transfer" risk; (5) Polyakova v8 production iteration анал — adding altitude-reasoning to prompt → +73% FDR (DELTA §2.4 §3.6 reject).
+
+- **Уверенность:** Medium-Low (defer to E5)
+- **Область:** Multi-task learning, Student feature quality
+- **Baseline:** Student no aux seg loss (anchor)
+- **Метрика:** R@1 на University-1652 val (Student); ΔCKA на Student dense features vs Teacher (feature pollution metric)
+- **Threshold для опровержения (anchor wins):**
+  - ΔR@1(aux seg) ≤ -0.3pp (feature pollution confirmed)
+  - ΔCKA(student-teacher) > 0.1 (feature drift confirmed)
+- **Опровержение alternative (aux seg helps):**
+  - ΔR@1(aux seg) ≥ +0.5pp с λ_aux ≤ 0.1 → multi-task helps, reconsider
+  - PCGrad gradient surgery enables: ΔR@1 ≥ +1pp при λ_aux=0.3 → adopt
+- **Зависимости:** Student trained anchor baseline; PCGrad implementation (REVIEW §8.3)
+- **Ресурсы:** 4 ablation arms × 3 seeds = 12 runs × 6h = 72 GPU-h
+- **Эксперимент:** E5.aux_seg_loss_ablation (post-E1, conditional)
+
+---
+
+## §10. Связанный модульный синтез
+
+Создан / обновлён:
+- [[СИНТЕЗ_6_segmentation]] — **создан** (initial scan, focus Pair A provider + cached uint8 fusion patterns)
+- [[СИНТЕЗ_3_fusion]] — обновлён §11 history (Pair A semantic categorical embedding pattern)
+
+Запись §11 «История обновлений»:
+- 2026-05-06 · DELTA §2.6 Pair A · D1-D7 + 4 hypotheses H_pair_A_5..8 + cached tensors paradigm shift (Mona-LoRA-A deprecated, nn.Embedding categorical primary).
+
+---
+
+## §11. Backlog impact
+
+Добавлены 7 новых items в [[READING_BACKLOG]]:
+
+- **P1** (1, promote): LMVMamba (already #17, promote to HIGH для §2.6)
+- **P2** (3): F11 GGeM verify, F18 SegDINO verify, CoOp + CLIP-Adapter
+- **P3** (4): DenseCLIP, MaskCLIP++, VL-Weed, AnyUp
+
+(Stepping Out ICCV'25 already в general OVSS literature — low priority)
+
+---
+
+## §12. Cross-links
+
+### К другим модулям
+- **§2.3** Pair D fusion — shared bottleneck + 5-way orthogonality + L2 9-zone routing — [[DELTA_E1_pair_D_text_fusion]]
+- **§2.4** VLM prompts — terrain inference + Object Inventory whitelist — [[DELTA_VLM_prompts_L1L2L3]]
+- **§2.5** Text encoder — LRSCLIP warm-start (verify v1 vs DGTRS-CLIP v2 cosine drift) — [[DELTA_E1_text_encoder_F_variants]]
+- **§2.13** Augmentation — cached tensors frozen rule — [[DELTA_E0_aug_refinement]]
+
+### К master plans
+- [[MASTER_synthesis_cached_tensors]] §3 (revised Arch A pair A — anchor source)
+- [[../../00_master/ROADMAP_E0_E9_unified]] Phase 3 + E5 ablation
+
+### К historical reference
+- [[REVIEW_segmentation_pairA]] (FC-CLIP-style, superseded by cached tensors regime)
+
+### Hypothesis bookkeeping
+- 4 новые гипотезы H_pair_A_5..8 (см. §9)
+- H_fus_A_1..4 status updates: 2 superseded, 1 confirmed, 1 invalid
+
+---
+
+## §13. Acceptance criteria
+
+DELTA принимается, если:
+- [x] anchor cached tensors regime (nn.Embedding + LRSCLIP-PCA + conv encoder + GGeM + FiLM-A) сохранён (no structural shift)
+- [x] §0.8 12 категорий FiLM-альтернатив проверены — anchor (E.7) primary, F3-research-A (E.11 MoFiLM) добавлен как research-only
+- [x] D1-D7 refinements — procedural / decision / ablation plan
+- [x] Conflict 1 (REVIEW_segmentation_pairA vs MASTER) explicitly resolved
+- [x] Conflict 2 (notation 1→32 vs 64→64) explicitly noted, code review action D8
+- [x] P0 backlog = 0 (full confidence)
+- [x] 4 новые гипотезы H_pair_A_5..8 имеют explicit thresholds
+- [x] Cross-DELTA consistency (§2.3/§2.4/§2.5/§2.13 sync verified)
+- [x] H_fus_A_1..4 status updates documented
+
+→ **DELTA APPROVED**. Pending P1 review (LMVMamba promote — для §2.6 H_pair_A_8 evidence) — recommended но не блокирующее. Code review D8 (notation fix) — pre-E1 must.
+
+---
+
+#delta #pair-A #segmentation #film #cvgl #priority/high #task/experiment #cached-tensors

docs/02_references/03_segmentation/_version_chain/DELTA_pair_A_seg_revised_v2.md

@@ -0,0 +1,105 @@
+---
+type: delta
+status: active
+date: 2026-05-12
+parent: "[[../00_overall/SPEC_fusion_ACF_MERIDIAN_v3]]"
+supersedes: "[[DELTA_pair_A_seg_revised]]"
+related:
+  - "[[DELTA_pair_C_chm_sat_v2]]"
+  - "[[../00_overall/HYP_fusion_variants_v2]]"
+  - "[[CVGL/2_hypotesis/03_fusion/03_text/DELTA_E1_pair_D_text_fusion_v2]]"
+applicable_to:
+  - E1
+tags:
+  - delta
+  - fusion
+  - pair-a
+  - segmentation
+  - v2
+  - post-F84-F85
+  - visloc-captions-link
+phase: E1
+author: claude
+---
+
+# DELTA: Pair A — Segmentation (revised) v2
+
+## Changelog v1 → v2 (2026-05-12)
+
+### Major changes
+
+1. **17-class inventory direct link к visloc captions** (NEW v2)
+   - Source: `visloc_satellite_captions.md` — Pair D captions используют **same 17-class инвентарь** для coverage% disclosure (Para 1)
+   - Sync: Pair A segmentation classes → Pair D caption coverage statistics
+   - Implication: Pair A и Pair D semantically aligned at class level (orthogonality regularizer carefully tuned λ_⊥(A,D))
+
+2. **Cached SegEarth-OV3 17-class via nn.Embedding(17,64)** — preserved v1 anchor
+
+3. **Differential-modal aug compat (H_arch_A_6)**
+   - $|X_A - X_D|$ — segmentation vs text divergence (visual vs descriptive)
+   - Potentially noisy due to alignment — may need careful adaptive λ_⊥
+
+### Preserved (v1)
+
+- nn.Embedding(17,64) + LRSCLIP-PCA-64 warm-start
+- Light conv encoder 4 layers (no Mona-LoRA-A в cached regime)
+- GGeM + FiLM-A head
+
+---
+
+## §1. Pair A architecture v2
+
+```
+Input: Sat image [B, 3, 256, 256]
+   ↓ SegEarth-OV3 frozen (cached features)
+Seg map [B, 1, H, W] class indices ∈ {0...16}
+   ↓ nn.Embedding(17, 64) → [B, 64, H, W]
+   ↓ Conv encoder 4 layers (64 → 64 → 128 → 128 → 128)
+Feature map [B, 128, H/8, W/8]
+   ↓ GGeM pooling
+[B, 128]
+   ↓ FiLM-A head MLP (128 → 256 → (2×1024×5))
+γ_A, β_A for Teacher blocks 20-24
+```
+
+## §2. Caption alignment (NEW v2)
+
+**Critical sync point**: visloc captions Para 1 enumerates **same 17 classes** для coverage% disclosure.
+
+Example caption:
+> "...8 buildings clustered along riverbank edges, mostly gray-roofed structures occupying roughly 30%, alongside dense green forest covering nearly half..."
+
+Здесь "buildings", "forest", "riverbank" — semantically aligned с Pair A seg inventory.
+
+**Implication для orthogonality**:
+- Pair A features semantic info from raw pixel-level seg
+- Pair D features semantic info from text-level coverage statistics
+- High correlation expected — **λ_⊥(A,D) should be adaptive** (allow partial redundancy)
+
+## §3. v2 research arms
+
+| Arm | Source | Activation |
+|:--|:--|:--|
+| **adaptive λ_⊥(A,D)** | NEW v2 caption-seg alignment | E1 default tuning |
+| **differential A-C aux** ($|X_A - X_C|$) | F85 DFE | E1 parallel |
+| **17-class coverage% as auxiliary scalar** | visloc Para 1 pattern | E1 research |
+
+## §4. Risks v2
+
+| Risk | Severity | Mitigation |
+|:--|:-:|:--|
+| **R1**: SegEarth-OV3 cached features stale | Low | Re-cache periodically |
+| **R2 NEW v2**: A-D high correlation breaks orthogonality | Medium | Adaptive λ_⊥(A,D); allow partial redundancy |
+| **R3**: 17 classes too coarse для some scenes | Low | Visible в captions Para 1 — handled via "minor elements" phrasing |
+
+## §5. Cross-references v2
+
+- [[../00_overall/SPEC_fusion_ACF_MERIDIAN_v3]]
+- [[../00_overall/DELTA_E1_teacher_5modal_Arch_A_final_v2]]
+- [[CVGL/2_hypotesis/03_fusion/03_text/DELTA_E1_pair_D_text_fusion_v2]] — **paired caption format**
+
+---
+
+[[DELTA_pair_A_seg_revised]] (v1) → archive. v2 supersedes (2026-05-12).
+
+#delta #fusion #pair-a #segmentation #v2 #post-F84-F85 #visloc-captions-link