proofs/lean/Consensus/Safety.lean

/-
  Consensus Safety Proof

  Theorem: No two honest nodes finalize conflicting values at the same height.

  This models the Lux consensus core as a state machine:
  - Nodes hold preferences and confidence counters
  - Each round: sample k peers, threshold vote, update confidence
  - Finalization: confidence reaches beta consecutive rounds

  The proof follows the standard metastable BFT argument:
  1. If honest majority prefers v, alpha threshold ensures v wins the round
  2. Consecutive wins accumulate confidence toward beta
  3. Once finalized, no honest node can switch preference (confidence monotonicity)
  4. Therefore two honest nodes cannot finalize different values

  Maps to Go code:
  - core/consensus.go: Block interface (ID, Height, Verify, Accept, Reject)
  - protocol/wave/wave.go: Wave.Tick (sampling, threshold, confidence)
  - protocol/wave/fpc/fpc.go: FPC threshold selection
  - core/types/: Decision enum (Undecided, Commit, Skip)
-/

import Mathlib.Data.Finset.Basic
import Mathlib.Data.Nat.Defs
import Mathlib.Tactic

namespace Consensus

/-- A value identifier (models ids.ID = [32]byte in Go) -/
abbrev ValueId := Fin (2^256)

/-- A node identifier (models types.NodeID) -/
abbrev NodeId := Nat

/-- Decision outcome (models types.Decision) -/
inductive Decision where
  | undecided : Decision
  | commit    : Decision
  | skip      : Decision
  deriving DecidableEq, Repr

/-- Consensus parameters (models wave.Config) -/
structure Params where
  n     : Nat        -- total nodes
  f     : Nat        -- max Byzantine nodes
  k     : Nat        -- sample size per round (wave.Config.K)
  alpha : Nat        -- vote threshold (ceil(alpha_ratio * k))
  beta  : Nat        -- confidence threshold (wave.Config.Beta)
  hf    : f < n / 3  -- BFT assumption
  hk    : k ≤ n      -- cannot sample more than population
  ha    : alpha ≤ k   -- threshold cannot exceed sample
  hb    : 0 < beta    -- need at least 1 round for finality

/-- Per-node state (models wave.WaveState + wave.prefs) -/
structure NodeState where
  preference  : Option ValueId    -- current preference (wave.prefs[item])
  confidence  : Nat               -- consecutive confirming rounds (WaveState.Count)
  finalized   : Option ValueId    -- decided value, if any
  honest      : Bool              -- Byzantine or honest

/-- Global consensus state -/
structure ConsensusState (p : Params) where
  round  : Nat
  nodes  : Fin p.n → NodeState

variable {p : Params}

/-- A node has finalized a value -/
def hasFinalized (s : ConsensusState p) (i : Fin p.n) (v : ValueId) : Prop :=
  (s.nodes i).finalized = some v

/-- A node is honest -/
def isHonest (s : ConsensusState p) (i : Fin p.n) : Prop :=
  (s.nodes i).honest = true

/-- Count of honest nodes preferring value v -/
noncomputable def honestPreferring (s : ConsensusState p) (v : ValueId) : Nat :=
  Finset.card (Finset.filter
    (fun i => (s.nodes i).honest = true ∧ (s.nodes i).preference = some v)
    Finset.univ)

/-- Count of honest nodes in the system -/
noncomputable def honestCount (s : ConsensusState p) : Nat :=
  Finset.card (Finset.filter
    (fun i => (s.nodes i).honest = true)
    Finset.univ)

/-
  Key invariant: once a node finalizes, its preference never changes.
  This models the early-return in wave.go:98-102:
    if state.Decided { return }
-/
axiom finalized_preference_stable :
  ∀ (p : Params) (s s' : ConsensusState p) (i : Fin p.n) (v : ValueId),
    hasFinalized s i v → s'.round > s.round → hasFinalized s' i v

/-
  At most one value reaches the alpha threshold in a round.

  `honestPreferring` counts over `Finset.univ`, i.e. the whole validator
  set rather than the k-sample, so the hypothesis needed here is
  `2 * alpha > n`: two disjoint sets of honest supporters, each of size at
  least alpha, cannot both fit in n nodes. `Consensus.ThresholdProof`
  proves exactly that as `unique_threshold_proven`.

  Without it the statement is false. Take n = 1000, f = 100, k = 20,
  alpha = 14 — legal under all four Params constraints — with 900 honest
  nodes split 450/450 between v1 and v2. Then honestCount = 900 > 2f, and
  alpha = 14 > 2*k/3 = 13, and both values clear alpha, so the conclusion
  would give v1 = v2.

  The hypothesis is severe and honestly so: with alpha ≤ k ≤ n it forces
  k > n/2, which is to sample more than half the network every round. In
  the sub-sampled regime the protocol actually runs, safety is
  probabilistic, and no deterministic theorem of this shape holds.
-/
axiom unique_threshold :
  ∀ (p : Params) (s : ConsensusState p) (v1 v2 : ValueId) (round : Nat),
    honestCount s > 2 * p.f →
    p.alpha > 2 * p.k / 3 →
    2 * p.alpha > p.n →
    -- If both values achieved alpha threshold in the same round,
    -- they must be the same value
    honestPreferring s v1 ≥ p.alpha →
    honestPreferring s v2 ≥ p.alpha →
    v1 = v2

/-
  Finalization window overlap (pure math component):
  Two intervals of length beta within [0, R) must share at least one
  element when R < 2 * beta. This is the pigeonhole/interval argument.

  If node i's window starts at a and node j's at b, with both windows
  fitting in [0, R), then a + beta ≤ R and b + beta ≤ R.
  If the windows were disjoint: a + beta ≤ b or b + beta ≤ a,
  so max(a + beta, b + beta) ≤ R and min(a,b) + 2*beta ≤ R,
  contradicting R < 2 * beta (since min(a,b) ≥ 0).
-/
theorem finalization_windows_overlap_math
    (R beta a b : Nat)
    (hR : R < 2 * beta)
    (ha : a + beta ≤ R)
    (hb : b + beta ≤ R) :
    -- The windows [a, a+beta) and [b, b+beta) share a round
    ∃ r, a ≤ r ∧ r < a + beta ∧ b ≤ r ∧ r < b + beta := by
  by_cases hab : a ≤ b
  · -- a ≤ b: overlap element is b (since b < a + beta)
    refine ⟨b, hab, ?_, le_refl b, by omega⟩
    -- Need: b < a + beta. From hb: b + beta ≤ R < 2 * beta, so b < beta.
    -- From ha: a + beta ≤ R, and hR: R < 2 * beta, so a < beta.
    -- If b ≥ a + beta, then b + beta ≥ a + 2 * beta > a + R ≥ a + a + beta,
    -- but more directly: a + beta ≤ b → a + 2*beta ≤ b + beta ≤ R < 2*beta
    -- → a + 2*beta < 2*beta → a < 0, contradiction since a : Nat.
    omega
  · -- b < a: overlap element is a
    push_neg at hab
    refine ⟨a, le_refl a, by omega, ?_, ?_⟩
    · omega
    · omega

/-
  Finalization window overlap (protocol connection):
  If two nodes finalize, the protocol produces a round where both
  values held alpha-threshold support. This connects the math above
  to the consensus state model.

  This remains an axiom because it models protocol behavior:
  finalization implies beta consecutive rounds of alpha-threshold support,
  and by finalization_windows_overlap_math those windows share a round.
  The axiom bridges the gap between "finalized" (an opaque predicate
  on ConsensusState) and "had alpha support for beta rounds" (which
  requires a trace model not present in this formalization).
-/
axiom finalization_windows_overlap :
  ∀ (p : Params) (s : ConsensusState p)
    (i j : Fin p.n) (v1 v2 : ValueId),
    hasFinalized s i v1 → hasFinalized s j v2 →
    -- There exists a round where both values achieved alpha threshold
    ∃ (s_overlap : ConsensusState p),
      honestPreferring s_overlap v1 ≥ p.alpha ∧
      honestPreferring s_overlap v2 ≥ p.alpha ∧
      honestCount s_overlap = honestCount s

/-
  Main Safety Theorem

  If two honest nodes finalize values at the same height,
  those values must be identical.

  Derived from unique_threshold + finalization_windows_overlap:
  1. Both finalized → windows overlap in some round (axiom)
  2. In that round, both achieved alpha (from overlap)
  3. By unique_threshold, v1 = v2
-/
theorem safety
    (p : Params)
    (s : ConsensusState p)
    (i j : Fin p.n)
    (v1 v2 : ValueId)
    (_hi : isHonest s i)
    (_hj : isHonest s j)
    (hfi : hasFinalized s i v1)
    (hfj : hasFinalized s j v2)
    (hbft : honestCount s > 2 * p.f)
    (halpha : p.alpha > 2 * p.k / 3)
    (hmaj : 2 * p.alpha > p.n)
    : v1 = v2 := by
  obtain ⟨s_overlap, hov1, hov2, hcount⟩ :=
    finalization_windows_overlap p s i j v1 v2 hfi hfj
  exact unique_threshold p s_overlap v1 v2 s_overlap.round
    (hcount ▸ hbft) halpha hmaj hov1 hov2

end Consensus