Register a go-wire SetExternalHandler for the "block" command in
services/legacy/peer that calls MsgBlock.Bsvdecode directly against
the network reader, replacing the default ReadMessageWithEncodingN
path which allocates a fresh []byte the full size of the declared
payload before parsing.

Heap profile pulled from bsva-ovh-teranode-ttn-eu-4 during a 25.4 GiB
RSS peak on 2026-05-16 showed 2.86 GB of inuse heap held by
ReadMessageWithEncodingN's payload buffer alone, the second-largest
contributor after per-tx scratch in go-wire (probe bundle archived
at probe/eu4-rss-peak-2026-05-16). On multi-GB blocks that buffer is
allocated once, copied from once, and then released, but it dominates
peak inuse during the decode window and aggravates GC pressure.

Tradeoffs:

- Payload checksum is no longer verified for the "block" command.
  The external-handler API gives the handler only the payload reader
  (the header is consumed before dispatch), so the DoubleHash
  comparison the default path performs cannot run here. This matches
  the existing extended-format path which already skips checksum
  verification because the cost outweighs the benefit at multi-GB
  sizes.
- The third return value of ReadMessageWithEncodingN (the raw payload
  bytes) is now nil for block messages. Only consumer in the legacy
  path is peer_server.OnBlock -> bsvutil.NewBlockFromBlockAndBytes,
  which stores the slice but reads it only if block.Bytes() is later
  called; no legacy netsync code path calls Bytes() on a block coming
  in over the wire (verified by grep), and bsvutil falls back to
  re-serializing from MsgBlock if it ever is, so correctness is
  preserved.

The handler wraps the reader in an io.LimitedReader so a malformed
varint cannot read past the declared payload boundary and desync the
next ReadMessage call, and drains any unread suffix before returning
so the stream stays aligned regardless of which error path
Bsvdecode takes.

Tests added: roundtrip equality vs the buffered path, drain-on-error
boundary check, idempotent registration.

- Reframe checksum doc: not a tradeoff. Downstream HandleBlockDirect
  enforces PoW, merkle reconstruction, and per-tx parse — any wire-
  level corruption a DoubleHash would have caught fails one of those.
  Per Siggi review on bsv-blockchain#885.
- Strengthen registration test: assert wire.ReadMessageWithEncodingN
  takes the streaming path (raw == nil, block decodes correctly) post-
  registration, and survives repeat Register calls. Replaces the
  prior no-op idempotent test. Per Copilot review on bsv-blockchain#885.

- Detect undersized payload: io.Copy on a LimitedReader returns nil
  if the underlying reader EOFs before N reaches 0, so a successful
  Bsvdecode followed by a truncated stream would silently report
  success and desync the next ReadMessage. Check lr.N after drain
  and return an error when the declared payload was not fully
  consumed. Per Copilot review on bsv-blockchain#885.
- Add TestStreamingBlockHandler_ShortStreamReturnsError covering it.
- Drop incorrect TCP CRC claim from the rationale comment (TCP uses
  a 16-bit additive checksum, not a CRC). Reframe what is actually
  given up: the early-rejection signal, not integrity itself. Per
  Copilot review on bsv-blockchain#885.