dash(daemonless): emit the self-derived MN checkpoint from the MN-CKPT bridge (task #154)

[frstrtr]

What

Give the MN-CKPT bridge the one-shot --dump-mn-checkpoint seam that until now
lived only on the empty-DIP3 replay-fold consumer. This lets a fresh anchor be
bumped from the baked checkpoint via a short forward fold (bridge SEEDs the baked
2513000 checkpoint and forward-applies to H), instead of a from-DIP3 replay (~a day).

Reward-safe: no change to the fold / apply / consensus path — only a dump hook and
a serialization overload.

The change (3 parts, each mirroring existing code)

1. mn_checkpoint_dump.hpp — factor the reward-safety-critical payload/.inc
   assembly out of emit_mn_checkpoint_dump() into finish_mn_checkpoint_dump()
   (byte format unchanged, now single-sourced so the two dumpers cannot diverge),
   and add a bridge overload of emit_mn_checkpoint_dump() / write_mn_checkpoint_inc()
   that serializes MnStateMachine::entries() (std::map<uint256, MNState>) via a pure
   MNState -> ReplayMNState field-copy. Reuses emit_mn_record_line() and the
   self-verify-through-parse_mn_checkpoint() guard VERBATIM — a set that fails its
   own self-parse is never written.
2. mn_checkpoint_lane.hpp — add set_dump_hook(target_height, fn) + a one-shot
   maybe_fire_dump() fired from the same clean-fold event that feeds
   set_on_block_applied (both apply sites). It therefore only fires on a
   forward-contiguous, non-gapped, non-payee-desynced, non-empty fold. Exact analog of
   FoldReplayConsumer::set_dump_hook. Unset = no-op.
3. main_dash.cpp — in the cold MN-CKPT bridge arm (the g_replay_fold_prestate.empty()
   branch), install the hook when --dump-mn-checkpoint is set, mirroring the replay-fold
   install; mutually exclusive by the prestate predicate. Block hash comes from OUR
   PoW-validated header chain.

getmnlistdiff is payout-stripped and cannot back a checkpoint — the bridge's
apply_block (dashd BuildNewListFromBlock parity) is the source, so every field is
self-derived from chain, no dashd.

Test evidence

• Build: c2pool-dash, DASH + real-BLS recipe (C2POOL_DASH_BLS=ON), Release — clean.
• Run (--run --embedded-mainnet --coin-p2p-discover --coin-p2p-peers 16 --web-port 0 --dump-mn-checkpoint 2522504 <out>, no --replay-bulk, no --replay-fold-prestate):
  • --dump-mn-checkpoint ARMED on the MN-CKPT BRIDGE — the new install fires.
  • Bridge SEEDs baked-2513000 (entries=2974) and folds forward with
    diverged=0 and payee_ok == applied throughout (reward-safe).
• Oracle-verify against fleet mainnet dashd (protx list registered 1 2522504):
  pending the dump landing (see note).

Note on pacing

On the build host the fold to H=2522504 is pace-limited by coin-P2P peer
contention from two co-resident full-history self-derive runs saturating the same
peer pool: each periodic getmnlistd fold-point stalls until the next mainnet tip
re-drives it (this base predates the on-demand-mnlist re-ask, #1292). It advances
reward-safely and lands the anchor, but in hours rather than minutes on a contended
host — an environment constraint, not a defect in this change. On an uncontended host
(or the combined onion base with the reduced-mismatch payee logic) it bumps in minutes.

Base / landing

Stacked on integrator/dashd-cut-header-backfill-canserve so the diff is exactly this
one commit. Intended to ride into the combined onion (#1293), which carries the same
dump feature + bridge and where the fold-consumer set_dump_hook install already lives.

DRAFT — reward-path, operator review. Do not merge / undraft.