diff --git a/docs/research/2026-08-08-345-mechanism-contract.md b/docs/research/2026-08-08-345-mechanism-contract.md
new file mode 100644
index 00000000..69bd359f
--- /dev/null
+++ b/docs/research/2026-08-08-345-mechanism-contract.md
@@ -0,0 +1,60 @@
+# #345 mechanism-session contract — who eats angled input on the steep-slope approach?
+
+**Date:** 2026-08-08. **Order:** after #344 lands. **This is an
+instrumentation session, not a fix session** — the fix contract gets written
+from the capture, not before.
+
+## Evidence in hand (do not re-derive)
+
+- `uphill-45deg-capture.jsonl`: 838 uniform stuck ticks — request ~0.23 m/tick
+  at 45° to the face, **result byte-identical to input position**,
+  `collisionNormalValid=True` with **`collN=(0,0,1)`** (straight up),
+  `slidingNormal=(0,0,0)`, carried contact = the ~6.8° approach terrain
+  `(-0.083,0.083,0.993)`, `transient=0x3` (Contact|OnWalkable).
+- `uphill-slide-capture.log`: the steep face itself was reached only 18
+  times; every reach took branch2 cliff-slide with a healthy applied
+  constraint. The eater is UPSTREAM of the face.
+- `uphill-AB-precampaign.jsonl`: identical 0.0%-yield signature on the
+  pre-campaign binary — **pre-existing**, Campaign S exonerated.
+- Ruled out by measurement: the #331 sliding absorb (no latch), cliff-slide
+  degeneracy (#32 fixed, visibly healthy), all Campaign S landings.
+
+## The one probe this session builds
+
+A per-tick TRANSITION PHASE TRACE, `ACDREAM_DUMP_TRANSIT_FAIL=1`, firing ONLY
+when a resolve returns zero XY displacement against a nonzero XY request
+(the stuck-tick predicate — self-selecting, so a normal session prints
+nothing). For each such tick, one line each for:
+1. every `TransitionalInsert` attempt: attempt #, phase reached
+   (env/building/objects), each phase's TransitionState, and on Collided the
+   colliding polygon's plane normal + which BSP/terrain source produced it;
+2. every step-up entry and its step-down verdict (reuse/extend
+   `ACDREAM_DUMP_STEPUP`'s existing lines — they already print the input
+   normal and the landing verdict);
+3. every `ValidateWalkable` outcome on the tick: dist, waterDepth, which
+   branch (resting / below-push / CheckWalkable-fail), and BOTH flag guards
+   (`oi.Contact`, `sp.StepDown`) at its `SetCollisionNormal` site — the
+   `collN=(0,0,1)` fingerprint most plausibly comes from a
+   `SetCollisionNormal(groundNormal)` *somewhere*, and this names which.
+4. the final `AdjustOffset` input/output pair for the tick.
+
+House rules: lives in `PhysicsDiagnostics` (rule 5), self-selecting trigger,
+zero cost when off, mover id on every line
+(`feedback_probe_identity_attribution`).
+
+## The session
+
+Build with the probe → the user repeats the exact 45° protocol (~1 min) →
+read the trace → the failing phase + normal source IS the mechanism → write
+the fix contract from it, citing the retail counterpart
+(grep-named-first: `CTransition::transitional_insert` family, the step-down
+refusal path, `validate_walkable`'s collision-normal writes — all already
+mapped in this repo's pseudocode docs).
+
+## Known trap inventory for the eventual fix (from this week's do-not-retry ledger)
+
+The #331 absorb is retail-faithful — whatever the fix is, that
+characterization pin must stay green. The #32 setter split must not be
+re-latched. AD-65's snap arm is byte-anchored. The step-up chain's
+`stepDownHeight = oi.StepUpHeight` oddity is retail-verbatim (#338's
+research). Do not touch any of these on a theory.