Cardano Updates

Decouple bracketFetchClient and bracketKeepAlive

Replace the `dataSize :: bytes -> Word` field of `ProtocolSizeLimits`
with a `BearerBytes` class:

    class BearerBytes bytes where
        bearerBytesSize :: bytes -> Word

Default instances cover `BS.ByteString`, `BL.ByteString`, `[Char]` and
`AnyMessage msg` (for tests).  Driver/Limits gains a `BearerBytes bytes`
constraint on the public entry points (`runPeerWithLimits` and friends).
The seven `byteLimits*` codec helpers no longer take a
`(bytes -> Word)` argument — call sites drop the `(fromIntegral .
LBS.length)` boilerplate.  Two test predicates that previously used
`dataSize` now call `bearerBytesSize` directly.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Replace the positional definition of `Reception` with named fields
`receivalTimes` and `received`:

    data Reception a = MkReception
      { receivalTimes :: !(IntMap Time)
      , received      :: !a
      }

Positional pattern matches continue to work; the fields just make
record-style access available where it reads better.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Wrap `Channel m a` recv values in a new `Reception a = MkReception
!(IntMap Time) !a`, recording per-chunk arrival times.  In the mux
demuxer the chunk's arrival time is paired with its starting byte
offset; codec/driver layers either strip the time map (the codec sees
only bytes) or propagate it through `wrapMiniProtocolTrailing` at the
mini-protocol callback boundary.  Tests/demos/benches use `IntMap.empty`
when synthesising trailing.

This is the network half of the November Leios demo's arrival-time
plumbing; consumers (e.g. Praos block-arrival tracing on the consensus
side) will be wired up separately.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Exports of `BlockFetch.ClientRegistry` are preserved, so no consequences
for downstream libraries.

Replace the `dataSize :: bytes -> Word` field of `ProtocolSizeLimits`
with a `BearerBytes` class:

    class BearerBytes bytes where
        bearerBytesSize :: bytes -> Word

Default instances cover `BS.ByteString`, `BL.ByteString`, `[Char]` and
`AnyMessage msg` (for tests).  Driver/Limits gains a `BearerBytes bytes`
constraint on the public entry points (`runPeerWithLimits` and friends).
The seven `byteLimits*` codec helpers no longer take a
`(bytes -> Word)` argument — call sites drop the `(fromIntegral .
LBS.length)` boilerplate.  Two test predicates that previously used
`dataSize` now call `bearerBytesSize` directly.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Set @TCP_NOTSENT_LOWAT = 32 000@ (~1.5 slow-start congestion window) and
@TCP_NODELAY = 1@ on each socket bearer at construction.  This caps the
kernel-side unsent buffer so that egress backpressure surfaces to the
mux scheduler instead of being absorbed silently — important for the
Leios demo where we want timely round-tripping of EB announcements and
fetches.

`TCP_NOTSENT_LOWAT` is not yet exposed by the `network` package, so we
use the numeric constants from Linux's `<netinet/tcp.h>`
(`IPPROTO_TCP` = 6, `TCP_NOTSENT_LOWAT` = 25) directly.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Wrap `Channel m a` recv values in a new `Reception a = MkReception
!(IntMap Time) !a`, recording per-chunk arrival times.  In the mux
demuxer the chunk's arrival time is paired with its starting byte
offset; codec/driver layers either strip the time map (the codec sees
only bytes) or propagate it through `wrapMiniProtocolTrailing` at the
mini-protocol callback boundary.  Tests/demos/benches use `IntMap.empty`
when synthesising trailing.

This is the network half of the November Leios demo's arrival-time
plumbing; consumers (e.g. Praos block-arrival tracing on the consensus
side) will be wired up separately.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

Exports of `BlockFetch.ClientRegistry` are preserved, so no consequences
for downstream libraries.

FetchClientRegistry was split into two parts, as a consequence
`bracketFetchClient` is decoupled from `bracketKeepAlive`.

As a consequence `bracketFetchClient` is decoupled from
`bracketKeepAlive`.

Snapshot-merge: parents are (old leios-prototype tip, origin/main tip)
but tree equals origin/main byte-for-byte. The leios-prototype branch
ref moves forward without losing the historic prototype commits — they
remain reachable as ancestors via the merge's first parent.

The chunks of the recreation land as ordinary commits on top of this
one.

Snapshot-merge: parents are (old leios-prototype tip, origin/main tip)
but tree equals origin/main byte-for-byte. The leios-prototype branch
now sits on top of current upstream main while keeping all historic
commits reachable via the first-parent ancestry.

Original tip tagged locally as pre-rebase-orig.
See /home/ch1bo/.claude/plans/1-leios-should-not-transient-hummingbird.md.

Note: a plain 'git rebase origin/main' was tried first per a suggestion;
it hit immediate conflicts on files upstream-deleted-leios-modified, so
we fall back to the snapshot-merge approach. The 57 demo-tuning commits
remain reachable via the merge's first parent for manual cherry-pick or
direct-upstream-PR later.

Bump cardano-base for Leios voting