sprint-log-LL1: backfill W1/W3/W4 agent reports

Three of the four wave-1 agents (W1, W3, W4) completed their code
changes but couldn't write their own agent-W{N}.md report files due
to permission snapshots captured pre-expansion. Main thread backfilled
the reports from the agents' task-notification summaries + code diff
inspection.

W1: NarsInferenceType extended with Intervention + Counterfactual,
plus W1 proactively patched 2 downstream exhaustive-match sites
(nars/inference.rs, orchestration_impl.rs) bridging to Abduction.

W3: edge.rs Reserved5/Reserved6 slots renamed to Intervention/
Counterfactual; from_bits decoder updated.

W4: triplet_graph.rs ContextTag enum + CounterfactualSpoG struct +
intervene_on() method.

Workspace cargo check passes. Wave 2 (W2 nars-engine-dispatch + W5
tests-e2e) spawns next with the expanded permissions.

Author: claude

.claude/board/sprint-log-LL1/agents/agent-W1.md

@@ -0,0 +1,17 @@
+# Agent W1 — nars-variants (LL1)
+
+**Status:** COMPLETE
+**File:** `crates/lance-graph-planner/src/thinking/nars_dispatch.rs`
+**Build:** `cargo check -p lance-graph-planner` exits 0
+
+**Changes:**
+- Added `NarsInferenceType::Intervention` (Pearl rung 2 do-calculus, `confidence_modifier=0.85`) with full doc comment citing arXiv:2510.01539 and ICM invariance
+- Added `NarsInferenceType::Counterfactual` (Pearl rung 3, 3-step abduce→intervene→predict, `confidence_modifier=0.70`) with doc comment
+- Extended `route()` and `detect_from_query()` to handle both new variants
+- Added `confidence_modifier()` impl method
+
+**Side effects (W1 patched proactively — flag for meta-review):**
+- `crates/lance-graph-planner/src/nars/inference.rs` — bridged new variants to existing `NarsInference::Abduction` semiring (with comment noting W2 should extend further)
+- `crates/lance-graph-planner/src/orchestration_impl.rs` — same exhaustive-match patch
+
+**Notes:** Agent could not write this report file itself (permission snapshot pre-expansion); backfilled by main thread. AGENT_ORCHESTRATION_LOG line was written by W1 via `tee -a`.

.claude/board/sprint-log-LL1/agents/agent-W3.md

@@ -0,0 +1,13 @@
+# Agent W3 — causal-edge-g (LL1)
+
+**Status:** COMPLETE
+**File:** `crates/causal-edge/src/edge.rs`
+**Build:** Workspace `cargo check` exits 0
+
+**Changes:**
+- Reclaimed `Reserved5` slot → `InferenceType::Intervention` (do-calculus, Pearl rung 2)
+- Reclaimed `Reserved6` slot → `InferenceType::Counterfactual` (Pearl rung 3)
+- Updated `from_bits()` decoder accordingly
+- Doc comments cite Pearl's do-operator + reference `CausalEdge64::counterfactual_ready` for the confidence gate
+
+**Notes:** Agent's textual task-notification was off-task (permission-pattern audit) but its actual code modification landed correctly. Backfilled by main thread. No bit-layout change beyond renaming Reserved slots; binary compat preserved.

.claude/board/sprint-log-LL1/agents/agent-W4.md

@@ -0,0 +1,14 @@
+# Agent W4 — arigraph-intervene (LL1)
+
+**Status:** COMPLETE
+**File:** `crates/lance-graph/src/graph/arigraph/triplet_graph.rs`
+**Build:** Workspace `cargo check` exits 0
+
+**Changes:**
+- Added `pub enum ContextTag` with variants `Observation` (default) and `Intervention` (Pearl rung 2 marker)
+- Added `pub struct CounterfactualSpoG { triplet, context }` — caller-owned value representing the substituted triple
+- Added `TripletGraph::intervene_on(subject, predicate, new_object)` method producing a `CounterfactualSpoG` with `ContextTag::Intervention`
+- Original graph NOT mutated (read-only borrow on self)
+- Sentinel byte `0xFF` in `raw_g` field documented as a placeholder until contract-level G enum lands
+
+**Notes:** Agent completed silently (no explicit task-notification arrived but file diff confirms work). Backfilled report by main thread.