Fix ISOM-R2 SDPA dimension mismatch in streaming window

Fix ISOM-R2 SDPA dimension mismatch in streaming window

Fix ISOM-R2 SDPA dimension mismatch in streaming window

Fix ISOM-R2 SDPA dimension mismatch in streaming window

Fix ISOM-R2 SDPA dimension mismatch in streaming window

fix: add layers attribute and get_mask_sizes method across all cache classes for Transformers 4.50+ compatibility

fix: completely remove super().__init__ across custom Cache classes to eliminate call_super_init AttributeError on PyTorch 2.5+/Python 3.13

fix: bypass Cache.__init__ via torch.nn.Module.__init__ across all cache classes for complete compatibility with latest transformers on Colab

fix: version-safe Cache.__init__ cascade + strip ISOM-only kwargs from super().generate() for transformers 4.50+ compat

docs: AutoTokenizer+AutoModelForCausalLM quickstart, Top-K tuning guide

config: isom_r2_num_retrieved_chunks=4, add isom_r2_micro_window_size=512

feat: Top-K multi-micro-window retrieval - sinks+retrieved_pages+rolling_window active buffer, micro_window_size config param, AutoTokenizer/AutoModelForCausalLM drop-in support

docs: update README with empirical 1M real GitHub repository benchmark telemetry

benchmark: add 1M real GitHub repository benchmark telemetry suite

config: set isom_r2_max_context to 1048576 (1M context)

feat: upgrade ISOM-R2 engine to 1M context with DiskBackedPagePool & vectorized centroid retrieval

Fix SDPA shape mismatch: restrict vault/manifold injection to autoregressive decoding only

Add unitary multi-hop manifold anchor readout for 1M context decoding

Integrate 1M upgrades: NeedleVaultBuffer + ThreePathGate + omega_min into both cache paths

Release ISOM-R2-Coder-1.5B with 1,048,576 (1M) context architecture, Needle Vault, and sub-harmonic Lie floor

Upload official ISOM-R2-Coder-1.5B model weights, architecture, and documentation

initial commit