[inureyes]

Summary

examples/logit_trace left chunk 0 without the BOS anchor (no BOS prepend, and no context pass under PREFILL), so on a model trained behind a document separator the first window collapsed and pooled perplexity or disagreement figures read like a defect.

• Without PREFILL, chunk 0 gets the BOS prepended like every other chunk (target_offset = bos_prefix.len()).
• With PREFILL, a chunk with no history gets a context pass of the BOS alone, so the traced forward keeps the chunk's width and target_offset 0. An empty bos_prefix skips it, so the data rows are unchanged for tokenizers without a BOS.
• The corpus is still encoded without special tokens; adding them would shift every chunk boundary.
• A # bos_anchor metadata line records the rule. compare_logit_traces.py reads only the model key, so parsing is unaffected.
• scripts/laguna_oracle_trace.py plan_chunks and its docstring follow the same rule and write the same metadata line.
• docs/benchmarks.md ("Judging a change that moves the numbers") says older traces are not comparable at chunk 0; the Laguna entry in docs/supported-models.md carries the re-measured figures.

Validation (GB10, real checkpoints)

Checkpoint poolside/Laguna-XS-2.1-NVFP4 (downloaded with hf download to ~/models/mlx/laguna-xs-2.1-nvfp4), corpus docs/architecture.md (4016 tokens). The float32 oracle runs on this aarch64 CUDA host (torch 2.14.1+cu130, transformers 5.19.0 in a scratch venv, --device cuda).

Arm	chunk 0 ppl, width 256	chunk 1 ppl
oracle, before	364.3	22.25
oracle, after	19.93	22.25
mlxcel, before	378.2	22.16
mlxcel, after	19.86	22.16

• compare_logit_traces.py oracle mlxcel after the fix: width 256 (256 2 8 0), 20/512 top-1 disagreements, 0/250 on decided positions; width 8 behind 512 (8 80 8 512), 27/640, 0/317 decided. Both exit 0, so (chunk, pos, target) agree on every row.
• Only chunk 0 changes: every row of chunks 1+ is byte-identical between the pre-fix and fixed traces, in mlxcel at both shapes and in the oracle at width 256. Under PREFILL chunk 0 (8 rows at the corpus head) goes from perplexity 762279 to 229.3, against 228.2 in the oracle.
• qwen2.5-0.5b-instruct-4bit (empty bos_prefix): # bos_anchor none, and the data rows are byte-identical before and after at 256 2 8 0 and 8 4 8 64.
• cargo clippy --profile test-fast --features cuda --example logit_trace -- -D warnings and cargo fmt --all -- --check are clean. The scripts/ci contract checks pass.

Closes #1785