acdream/docs/research/2026-05-21-open-items-pickup-prompt.md

Open items pickup prompt — 2026-05-21 session aftermath

After the 2026-05-21 collision-fix session, six discrete items remain. This doc gives a fresh session the full landscape: how the items relate, what depends on what, what order makes sense.

The pasteable session-start prompt is at the bottom of this doc.

The landscape at a glance (updated 2026-05-20)

#	Item	Domain	Depends on	M2-blocker?
#90 ping-pong	NEW TOP BLOCKER — CellId flips between outdoor 0xA9B40022 and indoor vestibule 0xA9B40164 on each wall push-back at the inn doorway; user perceives "walls walk through everywhere"	Collision — cell tracking	—	YES (M2)
A4	Walls walk-through in vestibule cells — multi-cell BSP iteration	Collision	—	CLOSED 2026-05-20 (shipped, but dormant in practice until #90 is fixed)
stairs	Stairs walk-through — likely same root cause as #90 (also a doorway-edge geometry issue)	Collision	#90	YES (likely subsumed by #90)
A2	PHSP inversion — polygon_hits_sphere early-return logically inverted vs retail	Collision math	—	Low (subtle correctness only)
A3	Synthesis removal — delete TryFindIndoorWalkablePlane, rely on retail CP retention	Architectural	A4 + #90 + A2	No (cleanup, retail-faithful)
L-indoor	Lighting indoors broken	Rendering	—	No (M7 polish)
L-spotlight	Items projecting spotlight on walls	Rendering	—	No (M7 polish)

Two domains, one critical-path chain

The 6 items split cleanly:

Domain 1 — Collision (M2 critical path). A4, stairs, A2, A3. These block "kill a drudge" because dungeons have drudges and dungeons have walls/stairs/floors that need to behave correctly. The dependency chain is:

                  A4 (multi-cell iteration)
                       │
            ┌──────────┴──────────┐
            ▼                     ▼
        stairs (verify          A3 (remove
        post-A4)                synthesis,
                                relies on
                                A4 retention)
                                     ▲
                                     │
                                 A2 (PHSP fix
                                 — also useful
                                 standalone)

A4 is the biggest user-visible win and it unblocks A3. A2 is a small self-contained correctness fix that pairs naturally with A3. Stairs are likely an A4 side-effect.

Domain 2 — Rendering (M7 polish). L-indoor + L-spotlight. These don't affect gameplay correctness — the world just looks wrong. Different code paths (lighting, not physics), different files, different domain knowledge. Best tackled in their own session, ideally after collision is solid so the visual artifacts are easier to isolate.

Why this order

1. A4 first. Biggest user-visible improvement. Closes the "vestibule cells don't block walls" gap by querying every cell the sphere overlaps, not just the one cell the player's center is in. Retail oracle: CTransition::check_other_cells at acclient_2013_pseudo_c.txt:272717-272798. Existing CellTransit.FindCellList already enumerates the right cells; A4 wires that into FindEnvCollisions. Probably 1-2 days.

2. Verify stairs after A4. If A4 closes vestibule walls, it probably also closes stairs (same architectural gap — stairs span cells). If stairs still walk-through after A4, investigate per-cell physics-poly coverage for stair geometry as a separate sub-issue.

3. A2. One-line flip in PolygonHitsSpherePrecise (BSPQuery.cs:117) plus a unit test for the tangent boundary. Improves correctness across every BSP query (walls, step-up, step-down). Pairs cleanly with A3.

4. A3. Once A4 + A2 land, the architectural cleanup becomes safe. Delete TryFindIndoorWalkablePlane (TransitionTypes.cs:1294) and the synthesis call site. Retail's grounded path doesn't synthesize CP — it retains via Mechanisms A/B/C (already in our code at BSPQuery.cs:1615, TransitionTypes.cs:2618, TransitionTypes.cs:896). The 2026-05-20 session tried A3 prematurely (Bug A) and reverted because the doorway-exit case relied on multi-cell iteration that wasn't there. A4 closes that gap.

5. Lighting (M7). Separate session. Different domain. Defer until M2 ships.

Per-item starter notes

A4 — multi-cell BSP iteration

Read first:

• §"Phase A4" of docs/research/2026-05-21-collision-fixes-shipped-handoff.md
• acclient_2013_pseudo_c.txt:272717-272798 (check_other_cells)
• src/AcDream.Core/Physics/TransitionTypes.cs:1407-1559 (FindEnvCollisions — change site)
• src/AcDream.Core/Physics/CellTransit.cs (helpers already exist)

Approach:

• Extract a "cell_array" set from the player's current cell via the existing CellTransit BFS.
• Iterate each cell, run BSPQuery.FindCollisions against each one.
• Combine results: any cell returning Collided/Adjusted halts; any cell returning Slid is remembered; all OK = return OK.

Acceptance: Walk into the Holtburg inn vestibule (cell 0xA9B40164). Walls in cell 0xA9B40157 should now block when the player's sphere extends into them, even though the player's center is still in the vestibule.

Stairs walk-through

Strategy: verification-only, after A4. Launch with the same probe set, walk up the inn stairs, watch the [indoor-bsp] results. If stair hits fire correctly, done. If not, investigate the cell's physics-poly data — stairs may be packed as static objects rather than cell-structure polys.

A2 — PHSP inversion fix

Bug: BSPQuery.cs:117 has if (MathF.Abs(dist) > rad) return false; — bails when sphere is FAR from plane. Retail's polygon_hits_sphere_slow_but_sure at acclient_2013_pseudo_c.txt:322509-322517 does the opposite — bails when sphere is OVERLAPPING plane.

Fix: flip the comparison. New unit test for the tangent boundary (sphere center at radius distance from plane → continue, not reject).

Caveat: doesn't fix walkable synthesis on its own — AdjustSphereToPlane also rejects at the tangent boundary (strict < check on interp). The two together gate the synthesis. Fixing A2 alone changes which side of the boundary the rejection happens on but doesn't close the gap. Pair with A3 for the full benefit.

Read first:

• §"Phase A2" of docs/research/2026-05-21-collision-fixes-shipped-handoff.md
• docs/research/2026-05-21-walk-miss-capture-findings.md (the smoking-gun analysis of the 2 cm boundary)

A3 — synthesis removal

Bug: retail's grounded path doesn't re-synthesize ContactPlane. It retains via three mechanisms (Path 4 land, LKCP proximity restore, post-OK step-down probe — all already in our code). Our TryFindIndoorWalkablePlane runs every frame and is unfaithful.

Fix: delete TryFindIndoorWalkablePlane (TransitionTypes.cs:1294) and its call site. ~500 lines deleted.

Critical prerequisite: A4 must ship first. The 2026-05-20 session tried A3 prematurely (Bug A) and reverted because doorway transitions caused free-fall — Mechanism C couldn't find a floor poly at the threshold because the indoor cell's BSP had no coverage past the doorway, and multi-cell iteration wasn't there to query the adjacent cell.

Read first:

• docs/research/2026-05-20-indoor-walking-bug-a-handoff.md (Bug A's premise + reversion)
• §"Phase A3" of the 2026-05-21 handoff

L-indoor — lighting broken inside

Symptom: lights inside buildings don't illuminate correctly.

Likely areas:

• Cell-light association (which lights belong to which cell)
• Light visibility culling (visible-cells set + light bounds)
• Per-light projection matrix indoors

Domain: rendering, not physics. Separate session.

L-spotlight — items projecting spotlight on walls

Symptom: held items (torches, etc.) project spotlight effects onto walls in unexpected directions.

Likely areas:

• Per-entity light direction transform
• LightingHookSink owner-tracking

Domain: rendering, not physics. Separate session.

CLAUDE.md rules to remember

1. Work-order autonomy. You pick what to work on. Recommended order above but adjust if you find something blocking.
2. No workarounds, retail-faithful. Same rule that drove A1 through A1.7. If a fix starts to look like a band-aid, stop.
3. Probe-first, design-second. Already have rich probes ([indoor-bsp], [cell-transit], [cell-cache], [walk-miss], [floor-polys], [resolve-bldg]). Capture before theorizing.
4. Visual verification is the acceptance test. Walk the building after each fix.
5. Stop signals. Three failed visual verifications in a session = write a handoff, don't push for a fourth.
6. Don't enable ACDREAM_PROBE_RESOLVE for live play. It lagged the client last session (400k+ lines at 30 Hz).
7. Subagent policy: Sonnet for implementers, Opus only for load-bearing review.
8. Worktrees. Use the superpowers:using-git-worktrees skill to create a fresh worktree branched from main before touching code.

Launch command (light probes only)

$env:ACDREAM_DAT_DIR              = "$env:USERPROFILE\Documents\Asheron's Call"
$env:ACDREAM_LIVE                 = "1"
$env:ACDREAM_TEST_HOST            = "127.0.0.1"
$env:ACDREAM_TEST_PORT            = "9000"
$env:ACDREAM_TEST_USER            = "testaccount"
$env:ACDREAM_TEST_PASS            = "testpassword"
$env:ACDREAM_DEVTOOLS             = "1"
$env:ACDREAM_PROBE_INDOOR_BSP     = "1"
$env:ACDREAM_PROBE_CELL           = "1"
$env:ACDREAM_PROBE_CELL_CACHE     = "1"
dotnet build -c Debug
dotnet run --project src\AcDream.App\AcDream.App.csproj --no-build -c Debug 2>&1 |
    Tee-Object -FilePath "launch.log"

UTF-16LE → UTF-8 conversion for grep:

Get-Content launch.log -Encoding Unicode |
    Out-File launch.utf8.log -Encoding utf8

---

The pasteable session-start prompt

Open a new Claude Code session in the main acdream worktree (C:\Users\erikn\source\repos\acdream, branch main at SHA f80b537 or later). Then paste:

---

Pick up the acdream open-items cleanup. After the 2026-05-21 session,
6 items remain across collision + rendering.

1. Read docs/research/2026-05-21-open-items-pickup-prompt.md FIRST.
   It maps the 6 items, their dependencies, and the recommended
   order (A4 → verify-stairs → A2 → A3 → lighting in a separate
   session). Each item has its own "Read first" anchor list inside.

2. Branch state: main is at f80b537 with all 2026-05-21 fixes
   landed (A1, A1.5, A1.6, A1.7 + probe spike + handoff docs).
   Build green, 1129-test baseline holds, four user-visible
   improvements visually verified. Local main is ahead of
   origin/main (origin at 7034be9); push only if explicitly asked.

3. **Set up isolation FIRST.** Use the superpowers:using-git-worktrees
   skill to create a fresh worktree from main. Don't work directly
   in the parent worktree. The 2026-05-21 session's worktree
   (claude/lucid-goldberg-1ba520) is identical to main and can be
   removed.

4. Recommended first phase: **A4 (multi-cell BSP iteration)**. It
   has the biggest user-visible payoff (closes vestibule-cell wall
   walk-through, likely closes stairs too) and unblocks A3
   architectural cleanup. Retail oracle is at
   acclient_2013_pseudo_c.txt:272717-272798 (CTransition::check_other_cells).
   Existing CellTransit helpers (FindCellList, FindTransitCellsSphere,
   AddAllOutsideCells) already enumerate the right cells; A4's
   work is wiring them into Transition.FindEnvCollisions.

5. Use the superpowers:brainstorming skill before writing A4 code.
   A4 is a real architectural change (multi-day, 2 files modified +
   tests) and deserves its own spec + plan. Don't shortcut it.

6. CLAUDE.md rules:
   - No workarounds. Retail-faithful.
   - Probe-first, design-second.
   - Visual verification at Holtburg inn cell 0xA9B40164 vestibule
     is the A4 acceptance test (walls in adjacent cell 0xA9B40157
     should block when the player straddles the boundary).
   - Don't enable ACDREAM_PROBE_RESOLVE for live play (lags the
     client). Use [indoor-bsp] + [cell-transit] + [cell-cache] only.
   - Three failed visual verifications = handoff, not a fourth attempt.

7. M2 ("kill a drudge") is the active milestone. A4 + stair
   verification + A2 + A3 are all on the M2 critical path because
   dungeons need walkable indoor space. Lighting is M7 polish;
   defer.

8. Launch command (light probes only):
   $env:ACDREAM_DAT_DIR              = "$env:USERPROFILE\Documents\Asheron's Call"
   $env:ACDREAM_LIVE                 = "1"
   $env:ACDREAM_TEST_HOST            = "127.0.0.1"
   $env:ACDREAM_TEST_PORT            = "9000"
   $env:ACDREAM_TEST_USER            = "testaccount"
   $env:ACDREAM_TEST_PASS            = "testpassword"
   $env:ACDREAM_DEVTOOLS             = "1"
   $env:ACDREAM_PROBE_INDOOR_BSP     = "1"
   $env:ACDREAM_PROBE_CELL           = "1"
   $env:ACDREAM_PROBE_CELL_CACHE     = "1"
   dotnet build -c Debug
   dotnet run --project src\AcDream.App\AcDream.App.csproj --no-build -c Debug 2>&1 |
       Tee-Object -FilePath "launch.log"

State the milestone + chosen phase in your first action.