Add missing clean-pretrained gradient-fixed CR2 iteration 2k inference weights

.gitattributes

@@ -135,3 +135,5 @@ c3-3/cr4_fullprefix_gradctxfix/iter_000005000/model/.metadata filter=lfs diff=lf
 c3-3/cr4_fullprefix_gradctxfix/iter_000005000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
 c3-3/cr4_fullprefix_gradctxfix/iter_000010000/model/.metadata filter=lfs diff=lfs merge=lfs -text
 c3-3/cr4_fullprefix_gradctxfix/iter_000010000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
+c3-3/cr2_fullprefix_gradctxfix/iter_000002000/model/.metadata filter=lfs diff=lfs merge=lfs -text
+c3-3/cr2_fullprefix_gradctxfix/iter_000002000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text

c3-3/cr2_fullprefix_gradctxfix/iter_000002000/README.md

@@ -0,0 +1,76 @@
+# CR2 clean-pretrained, full-prefix gradient-fixed compressor: iteration 2000
+
+This is the complete **2k** CR2 inference checkpoint from the current c3-3
+clean-pretrained rate sweep. It is not the k18 smoke-test model, an older
+warm-start model, or a c1-1 checkpoint. Training-run identifier: `fullprefix_memoryonly_cr2_c3_3_gradctxfix_gbs8_lr1e4_stagedcosine_clip5_raw_2k`.
+The independently checked trainer record is iteration 2000.
+
+## Weights and code
+
+Keep the complete `model/` directory, including hidden `.metadata` and all four
+`.distcp` files (three are empty original shards). Optimizer/trainer state is
+not included: this is an inference package, not a fully resumable training save.
+`sha256_manifest.json` records the exact size and SHA256 of every model file.
+
+GitHub: https://github.com/a-little-hoof/video-compress
+Published inference revision: `0c014000ef80e766bce385b947c2a3d34f155e2b`
+Branch: `exp/c3-strict-shared-history-20260911`
+
+This is an inference-code pin, not a claim about the historical training commit.
+GitHub and this private Hugging Face repository require separately authorized
+access. The pinned revision above supports the CR2 method shown below.
+
+Loader: `build_compressor` and `load_compressor_weights` in
+`scripts/context_compression/generate_compressed_context.py`.
+Architecture: Wan 1.3B, dimension 1536, FFN 8960, 12 heads, 30 layers,
+memory ratio 0.5, max-memory-tokens-per-frame 1024, packed temporal grid.
+Training is full-prefix, memory-only; raw anchors were not retained as generator
+conditioning alongside the compressed training prefix.
+
+## Explicit matched-interval M=1 example
+
+Use with `scripts/context_compression/generate_mbencht_c3_3_continuations.py`:
+
+```text
+--packed_compressor_ckpt /path/to/iter_000002000
+--checkpoint_iteration 2000
+--model_size 1.3B
+--memory_ratio 0.5
+--max_memory_tokens_per_frame 1024
+--methods c3_relax_slot_h1_cr2_matched_interval_replacement
+```
+
+The checkpoint path is the parent of `model/`. Raw bootstrap covers [0,42);
+compressed history begins at latent frame 42 and refreshes every 21 latents.
+The full compressed archive is contextualized before selection. Selectable
+complete memory units are restricted to the raw control's eligible history
+interval by whole-unit source-range containment. Source labels are not strict
+information boundaries after contextualization. Selection is legacy per-layer.
+M=1 uses one 1560-token frame-equivalent compressed history unit per layer at
+832x480, plus two raw sink frames and one raw recent frame.
+
+Retrieval policy is a runtime choice, not inherent in the checkpoint. Match
+policy, active memory budget, bootstrap and cadence across compression rates.
+In particular, the previously supplied CR4 full-archive example is a different
+selector; do not call that mixed configuration a compression-rate-only ablation.
+Matching the eligible interval does not establish full equivalence to the
+released Relax Forcing inference recipe. No new evaluation results are claimed.
+
+## Separate frozen generator and assets
+
+Generator repository: `worstcoder/rcm-Wan`.
+Revision: `142d77a331f1516fea772607c1bf25a2785a2505`.
+File: `Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step4.pt`.
+SHA256: `cc53e27e4501f58f43d1ca01dc2e5dde4df946ce5c85ab79695a902d929df008`.
+Do not substitute the TF-sCM-init variant.
+
+VAE/text encoder: `Wan2.1_VAE.pth` and `models_t5_umt5-xxl-enc-bf16.pth` from
+`worstcoder/Wan`, revision `18554de6b48df8ab2ae274972a076b27ad5cb602`.
+Use four sampling steps, sigma_max 1600, mid_t 15/16, 5/6, 5/8.
+
+## Current rate-sweep completeness
+
+The available non-smoke rates at 2k are CR2, CR4, CR6, CR8 and CR10.
+CR4/6/8/10 were already uploaded; this package adds the only missing rate in
+that current family. Historical c1-1, anchor-exclusive and warm-start variants
+are distinct experiments, not additional rates in this matched training family.

c3-3/cr2_fullprefix_gradctxfix/iter_000002000/inference_config.json

@@ -0,0 +1,21 @@
+{
+  "checkpoint_iteration": 2000,
+  "checkpoint_layout": "pytorch_distributed_checkpoint",
+  "fps": 16,
+  "generation_protocol": "official_continuous_seed0_v1",
+  "history_selection_protocol": "matched_interval",
+  "max_memory_tokens_per_frame": 1024,
+  "memory_ratio": 0.5,
+  "method": "c3_relax_slot_h1_cr2_matched_interval_replacement",
+  "mid_t": [
+    "15/16",
+    "5/6",
+    "5/8"
+  ],
+  "model_size": "1.3B",
+  "num_steps": 4,
+  "seed": 0,
+  "selection_policy": "legacy_per_layer",
+  "sigma_max": 1600,
+  "source_commit": "0c014000ef80e766bce385b947c2a3d34f155e2b"
+}

c3-3/cr2_fullprefix_gradctxfix/iter_000002000/model/.metadata

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43ddbab4c51ecebb070e1e89af03987f5a246b1f05ea700f6d94075fc725cfdb
+size 481482

c3-3/cr2_fullprefix_gradctxfix/iter_000002000/model/__0_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88f65aec6607374f0965a02ded2919d0401fb61bb7082fb1e8320792c2ba0ed7
+size 5681533593

c3-3/cr2_fullprefix_gradctxfix/iter_000002000/sha256_manifest.json

@@ -0,0 +1,36 @@
+{
+  "compression_rate": 2,
+  "files": [
+    {
+      "path": "model/.metadata",
+      "sha256": "43ddbab4c51ecebb070e1e89af03987f5a246b1f05ea700f6d94075fc725cfdb",
+      "size": 481482
+    },
+    {
+      "path": "model/__0_0.distcp",
+      "sha256": "88f65aec6607374f0965a02ded2919d0401fb61bb7082fb1e8320792c2ba0ed7",
+      "size": 5681533593
+    },
+    {
+      "path": "model/__1_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__2_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__3_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    }
+  ],
+  "inference_only": true,
+  "inference_source_commit": "0c014000ef80e766bce385b947c2a3d34f155e2b",
+  "iteration": 2000,
+  "memory_ratio": 0.5,
+  "schema": "c3-checkpoint-handoff/v2",
+  "training_run": "fullprefix_memoryonly_cr2_c3_3_gradctxfix_gbs8_lr1e4_stagedcosine_clip5_raw_2k"
+}