Improve storage performance with read pipelining, run lookup, request merging, and counters

doc/src/driver/storage/driver.storage.mgr.md

@@ -219,3 +219,37 @@ Drop-in procedure replacements for callers using the old void-returning API. For
 - `dispatch_next` is an internal procedure called with interrupts disabled. It dequeues the head of the CFIFO, allocates a heap copy via `ioreq_copy` (for ISR safety), and invokes the appropriate `dispatchRead` or `dispatchWrite` on the device. If neither dispatch pointer is set, the request is failed with `eDeviceNotFound`.
 - The blocking path relies on pointer-dereferenced spin (`puint32(@wait.Done)^`) to prevent FPC from caching the `Done` flag in a register across the `hlt` instruction.
 - Devices without `dispatchRead`/`dispatchWrite` set (i.e., drivers not yet migrated to Phase 2) will have `submit_io` return `eDeviceNotFound`. Legacy drivers must be migrated before they can use the new I/O path.
+
+
+## Bounded performance path
+
+The manager merges up to 32 consecutive queued requests when operation, device,
+LBA range, and virtual buffer range are contiguous. It never searches past the
+queue head or delays a lone request to create a merge. Drivers opt in through
+`mergeMaxSectors`; SATA AHCI advertises 65,536 sectors, while other drivers default
+to zero (disabled). A merged command delivers one completion to each original
+request, in FIFO order. A command error fails every member conservatively.
+
+`storage_get_stats` takes an interrupt-protected snapshot of the device counters.
+`DISK info <index>` displays accepted/completed/failed logical requests, hardware
+commands, merged requests, queue-full rejections, queue and active high-water
+marks, successful read/write sectors, request-size buckets, queue/service ticks,
+and FAT metadata cache hits/misses. Queue ticks are accumulated per logical
+request; service ticks are accumulated per hardware command. Tick values use
+`Counters.c32`, not milliseconds. Counters are unsigned 32-bit values, start at
+device registration, and wrap; use snapshot differences over bounded benchmark
+intervals. Size buckets are <=8, 9..256, and >256 sectors.
+
+FAT32 blocking and async reads share a pooled transfer engine with batches of up
+to four outstanding reads, capped by device queue depth. Destination slices are
+disjoint. On any error, the engine drains every accepted request before releasing
+its context and reports only the contiguous successful byte prefix. Unaligned
+edges and odd DMA addresses use the context's private scratch sector. Writes
+remain sequential. Run maps store cumulative file-cluster offsets; an open-file
+cursor handles sequential locality and binary search handles random seeks.
+
+`STORPERF` runs memory-backed regression checks (also included in boot unit tests)
+for reverse completion order, failures, unaligned reads, run lookup, queue
+rejection, merging limits, and callback accounting. It does not access real disks
+or measure hardware throughput. Compare counters alongside `STORBENCH` and file
+benchmarks on contiguous and fragmented files for actual performance validation.