diff --git a/docs/research/2026-07-05-m15-dungeon-pickup-prompt.md b/docs/research/2026-07-05-m15-dungeon-pickup-prompt.md
new file mode 100644
index 00000000..e01bcf6f
--- /dev/null
+++ b/docs/research/2026-07-05-m15-dungeon-pickup-prompt.md
@@ -0,0 +1,56 @@
+# Pickup prompt — M1.5 dungeon track: #137 collision first (paste into a fresh session)
+
+Read `claude-memory/project_physics_collision_digest.md` FIRST (the collision
+SSOT + DO-NOT-RETRY table — binding), then the ISSUES entries for **#137**,
+**#153**, **#138** (in that order), plus the prior-art closed fixes **#150**
+(ethereal door step-down skip), **#151** (portal-less wall shells), **#152**
+(building-cache re-base). Then work the track below.
+
+**Context:** the R5 movement-manager arc is DONE and merged — `main` ==
+`claude/vigorous-joliot-f0c3ad` @ `aa734f34` (facade close-out:
+`docs/research/2026-07-03-r5-managers/r5-wiring-handoff.md`). M1.5's remaining
+critical path is the dungeon track: **#137 dungeon collision → #153/#138
+teleport-out residuals → A7 dungeon lighting** (A7 last — its 5-dungeon-site
+capture protocol needs dungeons walkable first).
+
+**Slice 1 — #137 (dungeon collision at doors + wall openings), oracle-first:**
+1. AUTONOMOUS PREP: wire nothing yet. Re-read the EnvCell collision path
+   (`CellTransit`, the door apparatus, `ShadowObjectRegistry` per-cell
+   registration — A6.P4 architecture) and check whether the #150 ethereal
+   step-down skip (`TransitionTypes.cs`, retail pc:276795-276806) already
+   covers EnvCell dungeon doors or only building doors. Grep named-retail
+   before any fresh decompilation.
+2. REPRO (needs the user in-world): the 0x0007 dungeon, with
+   `ACDREAM_PROBE_RESOLVE=1` (+ `ACDREAM_CAPTURE_RESOLVE=<path>` if a replay
+   diff is warranted). Characterize per site: which door / which opening,
+   expected vs actual, open-vs-closed state. The issue text says symptoms are
+   NOT yet characterized — do not fix ahead of the repro.
+3. Fix per the digest workflow (retail decomp is the oracle; ACE second).
+   Acceptance: doors block/pass per open/closed state, wall openings pass,
+   solid walls block — matching retail in 0x0007. Register row per deviation,
+   same commit.
+
+**Slice 2 — #153 (portals only work once / teleport-onto-unstreamed-edge
+runaway), apparatus-first:** the issue is REOPENED with a narrowed residual
+(arrival on a NOT-yet-streamed landblock NEAR AN EDGE → cell-march + Z
+free-fall + garbage outbound cell). The issue text itself says **DO NOT
+guess-patch** — the original #145 burned 5 attempts. Build the capture +
+anchor/guard diagnostic at the crossing first. ⚠️ The issue's "likely fix
+shape" is a streaming-gap HOLD — but the user REVERTED a previous teleport
+hold as "shaky and bandaid" (`feedback_no_holds_for_slow_foundation`). If the
+evidence really does point at a hold (as retail's synchronous-load
+equivalent), STOP and get explicit user approval with the retail anchor in
+hand; do not ship it on your own authority.
+
+**Then:** the #138 acceptance run (portal out → world + objects + avatar all
+present across REPEATED round-trips, collision working — its component fixes
+are shipped; #151/#152 closed the wall residuals; what's left is the gate),
+and A7 lighting (#154 dungeon-dim, #142 windowed interiors, #143 portal
+swirl, #140 Fix D) as the M1.5 closer.
+
+**Gotchas:** visual gates + repro sessions need the user (plan autonomous
+decomp/probe prep around their availability); PowerShell Tee logs are UTF-16
+(`tr -d '\000'` before grep); the user manages client lifecycle (plain
+`dotnet run`, graceful close, ask if a rebuild is locked); probes:
+`ACDREAM_PROBE_RESOLVE` / `ACDREAM_PROBE_CELL` / `ACDREAM_CAPTURE_RESOLVE` /
+`ACDREAM_PROBE_ENT` are all wired and DebugPanel-toggleable.