🚬 Add bounded MTP + N-gram pipelining for latency-limited split inference

SKIPPY_PROTOCOL_TODO.md

@@ -155,28 +155,23 @@ chunk and every normal decode token.
 
 ## E. State Lifecycle Efficiency
 
-- [ ] Reduce checkpoint and restore round trips for speculative decode.
-  - Avoid checkpointing every `VerifySpan` when a cheaper journal or suffix
-    trim can preserve correctness.
-  - Use `SKIP_VERIFY_CHECKPOINT` only when the repair path is proven safe.
-
-- [ ] Add speculative journals.
-  - Record enough per-stage state to commit or discard a verify span without a
-    full restore where possible.
-
-- [ ] Improve suffix trim semantics.
-  - Make rejected speculative suffix rollback cheaper than full session restore.
-  - Validate recurrent-state families separately from transformer-only KV.
+- [x] Make speculative positions authoritative in stage-state v9.
+  - Decode and `VerifyWindow` messages carry the absolute position each stage
+    must have before execution.
+  - Stages ahead of that position rewind attention KV locally.
+  - Speculation never checkpoints, restores, trims by control message, or
+    replays a rejected prefix.
+  - Recurrent-state stages reject positional speculation instead of falling
+    back to the removed checkpoint protocol.
 
 - [ ] Add checkpoint hierarchy.
-  - Keep coarse prompt checkpoints and lightweight decode-span checkpoints.
+  - Keep coarse prompt checkpoints for non-speculative session/cache workflows.
   - Evict checkpoint state with explicit memory accounting.
 
 - [ ] Track state-control latency separately from activation latency.
   - Emit checkpoint, restore, trim, prefix-restore, and decode-fuse timings per
     stage.
-  - Use these metrics to decide whether direct final replies or journaled
-    rollback should land first.
+  - Keep speculative position-rewind counts and trimmed-token totals separate.
 
 ## Validation Gates
 

crates/llama-spec-bench/README.md

@@ -5,7 +5,7 @@ Local target/draft speculative decoding checker and benchmark.
 `llama-spec-bench` compares a target GGUF model with a draft GGUF model on a
 prompt set. It checks tokenizer compatibility, verifies speculative output
 against baseline target decoding, measures acceptance behavior, and reports
-projected verification costs.
+the measured target and draft costs of the serial diagnostic loop.
 
 ## Architecture Role
 
@@ -46,7 +46,7 @@ sequenceDiagram
         D-->>B: proposed tokens
         B->>T: verify proposals
         T-->>B: accepted prefix or rejection
-        B->>B: update acceptance and projection stats
+        B->>B: update acceptance stats
     end
     B->>B: compare against baseline target tokens
 ```
@@ -79,7 +79,8 @@ speculative output mismatch makes the command fail.
 - accepted/rejected draft token counts
 - acceptance rate and mean accepted tokens per window
 - baseline and speculative decode timing
-- projected rollback and scratch verification costs
+- measured target verification and draft proposal costs
+- measured speculative throughput and speedup versus target-only decoding
 - per-prompt text previews and mismatch index
 
 The default corpus path is