store: batch create.loc window preads and SIMD prefix

CHANGELOG.md

@@ -87,6 +87,11 @@ before 1.0).
   (packed pin outs). Write does not pread `create.loc`. Missing stamp is
   `Corrupt`.
 
+- **`create.loc` leftover stamp:** lookup reads/sums only through the highest
+  fk in each 1024-create window, preads those windows as one bulk batch (held
+  head-resolve session or `pread_batch`), and prefix-sums non-overflow windows
+  with SIMD (SSE2 on x86_64, NEON on aarch64). No cross-window loc cache.
+
 - **`store_reorg` overnight ASan OOM:** sibling ops no-op at 16 parked
   `held_bodies` and the tiny hub is not reopened. Recycle-every-16 grew
   libFuzzer RSS to the 2048 MiB cap (~45 min into the Sunday 1h job)

SCHEMA.md

@@ -370,7 +370,11 @@ inwit.loc / inwit.off / inwit.loc.ovf   # u16 strides; cold with inwit.body
 Loc byte **0** = overflow (`n_out ≥ 256` or txout aligned length ≥ 2048).
 `inwit.loc` **0** = overflow (`strides ≥ 65536`, i.e. ≥ 512 KiB). Missing ovf
 when a sentinel is set is `Corrupt("invariant: create.loc overflow missing")`
-(or inwit). Checkpoints (~22 MiB) are RAM; do not L2 `create.loc`.
+(or inwit). Checkpoints (~22 MiB) are RAM; do not L2 `create.loc`. Lookup `range_batch`
+reads and prefix-sums only through the highest fk in each 1024-create window
+(not the unused tail). Those window preads are **one** bulk batch (held
+head-resolve session, else `pread_batch`). Non-overflow windows use a SIMD
+prefix sum (`u8×8` SSE2 on x86_64, NEON on aarch64).
 
 One `create_loc_range_batch` yields both `(txout, spent)` and `n_out`. Lookup
 stamps both ranges; load copies the stamp; write appends loc and keeps the RAM