fix(network): block-sync stack review follow-ups

[evan-forbes]

Motivation

Review follow-ups ("flups") collected from the merged block-sync throughput stack
(#284, #286, #287, #288, #289). These were filed as deferred during stack review and
are bundled here as a single follow-up PR against feat/pre-release-main.

The two largest themes:

• Stop relying on fixed-cadence timers/poll loops where a real event already
  exists — two backstops remained after the stack landed (the fix(network): add block sync congestion control #288 merge kept both
  the demand-driven floor-funding path and the old timer backstop).
• Remove a config double-negative before release (disable_vct_fast_sync).

Solution

• sequencer: drop the redundant timer-driven floor-starvation shed (F1, fix(network): prioritize block sync floor requests #287).
  The 500ms FLOOR_STARVATION_SHED_INTERVAL / shed_tick arm is deleted. Floor
  rescue is already fully demand-driven: inline after each accepted body, plus the
  synchronous SequencerControlInput::FundFloorReservation path a floor requester
  drives when it cannot reserve budget. The periodic backstop was a pure
  fixed-cadence poll on top of those.

• reactor: make the floor watchdog event-driven (F2). Instead of polling on a
  fixed floor_watchdog_tick, the reactor arms a sleep to the earliest
  outstanding floor-claim deadline (a new non-allocating
  PeerRegistry::earliest_outstanding_deadline_at), mirroring the per-peer
  routine's existing own-timeout arm. The cancellation logic (run_floor_watchdog)
  is unchanged; only when it fires moves from a 1s poll to the exact deadline.
  Removes the floor_watchdog_tick config knob (+ accessor, default, const, and
  both stored test configs).

• consensus/state: rename consensus.disable_vct_fast_sync →
  consensus.vct_fast_sync (F3, fix(network): tune block sync throughput defaults #289). Default flips from false to true
  (same behavior, no double negative). Inverted at every consumer: the state-config
  mirror, select_source_mode, finalized_state resume guard, the zebrad
  startup wiring, unit/prop tests, the VCT design doc, and CHANGELOG. No
  serde(alias) — this is a clean pre-release rename (the old name never shipped in
  a stable release).

• work_queue: assert the take_in_range_budgeted precondition (F4, perf(network): pack block sync ranges by size hint #284).
  Adds debug_assert!(max_count > 0 && low <= high, ...) so an invalid call is a
  loud caller bug in debug/test builds, while release still degrades to an empty
  Vec rather than panicking a live node.

• block_sync: reword refactor-historical comments (F5, perf(network): retain raw block bodies in reorder backlog #286). Comments framed
  as "ported from / verbatim from reactor.rs / matches the previous
  run_peer / used to …" required knowing the pre-refactor code and carried
  rot-prone line numbers. Reworded across the module to describe current behavior;
  genuine "why it's shaped this way" comments were kept.

Dropped after discussion (left as-is): bumping MAX_BS_INFLIGHT_REQUESTS and the
32k DEFAULT_BS_MAX_INFLIGHT default.

Tests

• cargo fmt --all -- --check — clean.
• cargo clippy -p zebra-network -p zebra-consensus -p zebra-state -p zebrad --all-targets -- -D warnings — clean.
• cargo nextest run -p zebra-network -p zebra-consensus -p zebra-state — 1149/1150
  pass. The one failure (peer::handshake::tests::mutual_p2p_v2_legacy_upgrade_forms_zakura_connection)
  is a pre-existing concurrency flake unrelated to this change — it passes 3/3 in
  isolation and touches handshake code, not block-sync/config/VCT.
• F-specific surfaces all pass: vct::tests::source_mode_precedence (rewritten for
  the inverted flag), the consensus/state default tests, the stored-config TOML
  parse/round-trip, and the block-sync shed_top_* / admission tests.
• Being validated IRL on a running node.

Specifications & References

Originating review threads: #284, #286, #287, #288, #289.

Follow-up Work

• peer_routine's OUTBOUND_FULL_POLL_INTERVAL (10ms re-check while a peer's
  outbound stream is full) is the one remaining poll left in place: converting it
  needs a session-level "outbound capacity freed" notification that may not exist
  yet, and the recv gating is subtle. Left untouched for safety and noted here.

AI Disclosure

• AI tools were used: Claude Code (Opus 4.8) — implemented the changes (timer
  removal, event-driven watchdog, config rename, comment sweep) and ran the local
  fmt/clippy/test validation.

PR Checklist

• The PR title follows conventional commits format: type(scope): description
• The PR follows the contribution guidelines.
• This change was discussed in an issue or with the team beforehand.
• The solution is tested.
• The documentation and changelogs are up to date.

[evan-forbes]

happy to revert this as well fwiw

Remove a config double-negative before release (disable_vct_fast_sync).

might cause issues with existing deployments

[evan-forbes]

fixing conflicts now