ContextLattice v3.18.0 - Continuity Has an Identity

Release posture: generally available. The deterministic semantic profile,
paid entitlement, workspace isolation, restart persistence, rollback,
full-matrix, and live-runtime gates all pass. The release remains exact-first,
explicitly governed, and independently reversible.

Agents do not lose the plot because intelligence disappeared. They lose it
because task, session, branch, worktree, and objective were treated as the same
thing.

v3.18.0 gives work a durable identity of its own.

One task, many lanes

• A task identity survives agent, account, model, session, branch, and worktree
  changes.
• Execution lanes remain explicit, so parallel work cannot masquerade as one
  process.
• Agent sessions can reuse stable work without collapsing lane boundaries.

Exact before semantic

• Explicit IDs, external task IDs, and exact normalized objectives always win.
• Semantic matching runs only after exact lookup misses.
• A strong semantic match is still advisory; ambiguity abstains and no semantic
  candidate auto-merges.
• Manual merge and split require explicit operator attribution plus a reason
  and create immutable receipts.

Objectives that remember time

• Typed objective transitions are append-only and actor-attributed.
• Dependencies, supersession, task, lane, session, decision, outcome, and
  checkpoint links form a longitudinal graph.
• As-of replay requires both event time and ledger-recorded time, reconstructing
  what was actually known then without backdated or future-event leakage.
• Indexed graph selection, traversal, and replay cap nodes, links, edges,
  transitions, and inspections, and mark every partial response explicitly.
• A 500 KB output contract compacts oversized graphs without stale counts or a
  false complete claim.

Decisions with receipts

• decision_change.v1 records before, after, evidence triggers, confidence
  delta, alternatives, actor, rationale, reason code, and verification.
• Each decision and its linked objective transition commit as one ledger entry,
  preventing half-recorded provenance after a crash.
• Objective and decision writes have durable idempotency keys, so a response
  lost after persistence can be retried without duplicating history.
• Decision history exposes count exactness, omissions, hard query ceilings, and
  opaque cursor pagination instead of silently returning a partial list.
• Hidden reasoning and chain-of-thought fields are rejected. The product stores
  accountable provenance, not private scratch work.

Built to distrust ambiguity

• The local continuity ledger is owner-only, bounded, append-only, fsynced by
  default, and hash-chained.
• Startup verifies every sequence and hash.
• A lifetime writer lock prevents multiple gateway processes from racing the
  same continuity ledger.
• Interrupted final writes recover only the uncommitted tail; tampering,
  malformed committed data, capacity exhaustion, or uncertain persistence
  still fails closed.
• Explicit lossless compaction canonicalizes the verified ledger without
  deleting history or changing its chain head.
• The Codex session-store doctor now separates topology warnings from actual
  permission failures and recommends repair only when an exact probe records
  literal OS-denial evidence.
• The CLI remains primary. HTTP is the integration fallback. No MCP tool-count
  increase, new daemon, scheduler, runner subprocess, or Python live dependency
  was added.