# Indoor Cell Rendering Fix — Phase 2 Implementation Plan

> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.

**Goal:** Surface WB's silent `PrepareEnvCellMeshData` failures via an exception-capturing continuation in `WbMeshAdapter`, identify the root cause for the 26 missing-completion cells, then implement the targeted fix that lands the indoor floor rendering.

**Architecture:** `WbMeshAdapter.IncrementRefCount` captures the `Task<ObjectMeshData?>` returned by WB's `PrepareMeshDataAsync` and attaches a `ContinueWith` that logs faulted-task exceptions + clean-null results for EnvCell IDs only. Gated by the existing `ProbeIndoorUploadEnabled` flag — zero cost when off. Component 3 (the actual fix) is data-driven: the captured exception type + message determines the surgical code change.

**Tech Stack:** C# .NET 10, Silk.NET OpenGL, WorldBuilder's `Chorizite.OpenGLSDLBackend.Lib.ObjectMeshManager`. xUnit for any unit tests.

**Spec:** [`docs/superpowers/specs/2026-05-19-phase2-indoor-cell-rendering-fix-design.md`](../specs/2026-05-19-phase2-indoor-cell-rendering-fix-design.md).
**Phase 1 capture:** [`docs/research/2026-05-19-indoor-cell-rendering-probe-capture.md`](../../research/2026-05-19-indoor-cell-rendering-probe-capture.md).

---

## File Structure

| File | Status | Responsibility |
|---|---|---|
| `src/AcDream.App/Rendering/Wb/WbMeshAdapter.cs` | MODIFY (Task 1) | Capture `prepTask` from `PrepareMeshDataAsync`. Attach a `ContinueWith` for EnvCell IDs that emits `[indoor-upload] FAILED` on faulted tasks and `[indoor-upload] NULL_RESULT` on clean-null returns. |
| `launch.log` (and the user's walk-through) | NEW (Task 2) | Captured probe output. Drives Component 3's fix shape. Not committed. |
| `docs/research/2026-05-19-indoor-cell-rendering-cause.md` | NEW (Task 3) | One-page report documenting the captured exception type(s) + the chosen fix shape. Becomes Phase 2's "design closure" doc. |
| TBD-by-data (Component 3) | MODIFY (Task 4) | Fix shape depends on captured cause. Likely candidates: `WbMeshAdapter.PopulateMetadata`, `CellMesh.Build`, a guard at the dat-access call site, or a small WB fork patch. |
| `docs/research/2026-05-19-indoor-cell-rendering-verification.md` | NEW (Task 5) | Post-fix verification record: previously-missing cells now emit `[indoor-upload] completed`, visual confirmation. |
| `docs/plans/2026-04-11-roadmap.md` | MODIFY (Task 6) | Roadmap update: Phase 2 shipped, link to spec + research notes. |

---

## Task 1: Add exception-surfacing continuation in `WbMeshAdapter`

**Files:**
- Modify: `src/AcDream.App/Rendering/Wb/WbMeshAdapter.cs`

- [ ] **Step 1: Add `using System.Linq;` and `using System.Threading.Tasks;` if missing**

Open `src/AcDream.App/Rendering/Wb/WbMeshAdapter.cs`. Verify both `using System.Linq;` and `using System.Threading.Tasks;` are present at the top. Add them if not.

- [ ] **Step 2: Replace the fire-and-forget call with a captured task + continuation**

Find the `IncrementRefCount` method (around line 116). The current block looks like:

```csharp
public void IncrementRefCount(ulong id)
{
    if (_isUninitialized || _meshManager is null) return;
    _meshManager.IncrementRefCount(id);

    if (_metadataPopulated.Add(id))
    {
        PopulateMetadata(id);

        // WB's IncrementRefCount alone only bumps a usage counter; it does
        // NOT trigger mesh loading. We must explicitly call PrepareMeshDataAsync
        // so the background workers actually decode the GfxObj. The result
        // auto-enqueues into _stagedMeshData (ObjectMeshManager line 510),
        // which Tick() drains onto the GPU. Until that completes,
        // TryGetRenderData(id) returns null and the dispatcher silently
        // skips the entity — standard streaming flicker.
        //
        // isSetup: false — acdream's MeshRefs already carry expanded
        // per-part GfxObj ids (0x01XXXXXX). WB's Setup-expansion path is
        // unused.
        _ = _meshManager.PrepareMeshDataAsync(id, isSetup: false);

        // [indoor-upload] requested probe — only for EnvCell ids.
        if (RenderingDiagnostics.IsEnvCellId(id) && RenderingDiagnostics.ProbeIndoorUploadEnabled)
        {
            _pendingEnvCellRequests.Add(id);
            Console.WriteLine($"[indoor-upload] requested cellId=0x{id:X8}");
        }
    }
}
```

Replace the `_metadataPopulated.Add(id)` block body with this exact content (note: the `_ = _meshManager.PrepareMeshDataAsync(...)` line becomes `var prepTask = ...` — capture the task instead of discarding it):

```csharp
        PopulateMetadata(id);

        // WB's IncrementRefCount alone only bumps a usage counter; it does
        // NOT trigger mesh loading. We must explicitly call PrepareMeshDataAsync
        // so the background workers actually decode the GfxObj. The result
        // auto-enqueues into _stagedMeshData (ObjectMeshManager line 510),
        // which Tick() drains onto the GPU. Until that completes,
        // TryGetRenderData(id) returns null and the dispatcher silently
        // skips the entity — standard streaming flicker.
        //
        // isSetup: false — acdream's MeshRefs already carry expanded
        // per-part GfxObj ids (0x01XXXXXX). WB's Setup-expansion path is
        // unused.
        var prepTask = _meshManager.PrepareMeshDataAsync(id, isSetup: false);

        // [indoor-upload] requested probe — only for EnvCell ids.
        if (RenderingDiagnostics.IsEnvCellId(id) && RenderingDiagnostics.ProbeIndoorUploadEnabled)
        {
            _pendingEnvCellRequests.Add(id);
            Console.WriteLine($"[indoor-upload] requested cellId=0x{id:X8}");

            // Phase 2 — surface what WB's catch block silently swallows.
            // ObjectMeshManager.PrepareMeshData has try/catch at line 589
            // that calls _logger.LogError on exceptions and returns null.
            // We construct ObjectMeshManager with NullLogger so the log
            // goes nowhere. This continuation captures the same data
            // (scoped to EnvCell ids only). Runs on ThreadPool; non-
            // blocking. Zero cost when probe is off.
            ulong cellId = id;
            _ = prepTask.ContinueWith(t =>
            {
                if (t.IsFaulted && t.Exception is not null)
                {
                    var ex = t.Exception.InnerException ?? t.Exception;
                    var stack = (ex.StackTrace ?? "").Split('\n')
                        .Take(3).Select(s => s.Trim()).Where(s => s.Length > 0);
                    Console.WriteLine(
                        $"[indoor-upload] FAILED cellId=0x{cellId:X8} " +
                        $"exception={ex.GetType().Name}: {ex.Message} " +
                        $"stack=[{string.Join(" | ", stack)}]");
                }
                else if (t.IsCompletedSuccessfully && t.Result is null)
                {
                    Console.WriteLine($"[indoor-upload] NULL_RESULT cellId=0x{cellId:X8}");
                }
            }, TaskScheduler.Default);
        }
```

- [ ] **Step 3: Build**

Run: `dotnet build src/AcDream.App/AcDream.App.csproj -c Debug`
Expected: 0 errors, 0 warnings (any new warnings about discarded tasks are fixed by the `_ = prepTask.ContinueWith(...)` assignment).

- [ ] **Step 4: Run tests (sanity)**

Run: `dotnet test tests/AcDream.Core.Tests/AcDream.Core.Tests.csproj --filter "FullyQualifiedName~Rendering" -c Debug --nologo --no-build`
Expected: All 130 Rendering tests still pass (the change doesn't touch any tested code path — `WbMeshAdapter.IncrementRefCount` isn't covered by unit tests).

- [ ] **Step 5: Commit**

```bash
git add src/AcDream.App/Rendering/Wb/WbMeshAdapter.cs
git commit -m "$(cat <<'EOF'
feat(wb): surface WB-swallowed exceptions for EnvCell upload failures

Phase 1 confirmed 26/123 Holtburg cells silently fail in WB's
PrepareEnvCellMeshData / PrepareMeshData. WB's catch block at
ObjectMeshManager.cs:589 calls _logger.LogError(ex, ...) — but we
construct ObjectMeshManager with NullLogger, so the log is dropped.

Capture the Task from PrepareMeshDataAsync (previously fire-and-forget)
and attach a ContinueWith that, for EnvCell ids only when the probe
is on, logs:

  [indoor-upload] FAILED cellId=0x... exception=<Type>: <Message>
                          stack=[<top 3 frames>]
  [indoor-upload] NULL_RESULT cellId=0x...

Runs on ThreadPool — non-blocking. Zero cost when ProbeIndoorUploadEnabled
is off. AggregateException is unwrapped to InnerException for readability.
Stack truncated to top 3 frames.

Next: capture procedure, identify cause, target the fix.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)"
```

---

## Task 2: Capture procedure — run client, identify cause

This task is operator-driven, not subagent-driven. The user (not a subagent) walks the client. Subagent role is limited to launching + analyzing the log.

**Files:**
- New: `launch.log` (transient — not committed)

- [ ] **Step 1: Full solution build (sanity)**

Run: `dotnet build AcDream.slnx -c Debug --nologo 2>&1 | tail -10`
Expected: `Build succeeded. 0 Error(s)`.

- [ ] **Step 2: Gracefully close any prior `AcDream.App` instance**

```powershell
$proc = Get-Process -Name AcDream.App -ErrorAction SilentlyContinue
if ($proc) {
    $proc | ForEach-Object { $_.CloseMainWindow() | Out-Null }
    $proc | ForEach-Object { if (-not $_.WaitForExit(5000)) { Stop-Process -Id $_.Id -Force } }
    Start-Sleep -Seconds 3
}
```

- [ ] **Step 3: Launch with `ACDREAM_PROBE_INDOOR_UPLOAD=1`**

```powershell
$env:ACDREAM_DAT_DIR             = "$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_UPLOAD = "1"
$logPath = "launch.log"
Remove-Item $logPath -ErrorAction SilentlyContinue
dotnet run --project src\AcDream.App\AcDream.App.csproj --no-build -c Debug 2>&1 | Tee-Object -FilePath $logPath
```

Run in background via `run_in_background: true`.

- [ ] **Step 4: User walks Holtburg**

User waits for the client to reach in-world (~8-12 s), then:
- Walks into Holtburg Inn (where the floor was missing in Phase 1).
- Walks into 2-3 other nearby buildings to capture varied failure causes.
- Closes the client window with the close button (graceful — NOT taskkill).

- [ ] **Step 5: Analyze the log**

```powershell
$lines = Get-Content launch.log | Where-Object { $_ -match '\[indoor-upload\] (FAILED|NULL_RESULT)' }
Write-Host "Total failure lines: $($lines.Count)"
Write-Host ""
Write-Host "=== Distinct exception types (FAILED) ==="
$lines | Where-Object { $_ -match '\[indoor-upload\] FAILED' } |
    ForEach-Object { if ($_ -match 'exception=(\w+):') { $matches[1] } } |
    Group-Object | Sort-Object Count -Descending | Format-Table -AutoSize

Write-Host "=== Distinct NULL_RESULT count ==="
($lines | Where-Object { $_ -match 'NULL_RESULT' }).Count

Write-Host ""
Write-Host "=== Sample FAILED lines ==="
$lines | Where-Object { $_ -match '\[indoor-upload\] FAILED' } | Select-Object -First 10
Write-Host ""
Write-Host "=== Sample NULL_RESULT lines ==="
$lines | Where-Object { $_ -match '\[indoor-upload\] NULL_RESULT' } | Select-Object -First 5
```

Verify the previously-failing cells (from Phase 1: `0xA9B40100`, `0xA9B40111`, `0xA9B40112`, etc.) now appear in either FAILED or NULL_RESULT.

If they DON'T appear:
- Confirm the probe flag is on (check `$env:ACDREAM_PROBE_INDOOR_UPLOAD` reads `"1"`).
- Confirm the user actually walked into the failing cells.
- Possible BUG: the continuation isn't firing — check Task 1's edits for typos.

---

## Task 3: Write the cause report

**Files:**
- Create: `docs/research/2026-05-19-indoor-cell-rendering-cause.md`

- [ ] **Step 1: Write the report based on Task 2's output**

Create the file with this structure (replace bracketed sections with captured data):

```markdown
# Indoor Cell Rendering — Phase 2 Cause Report

**Date:** 2026-05-19
**Predecessor:** Phase 1 capture confirmed H1 (silent failure in WB).
**Capture method:** Task 1's `ContinueWith` surfaced WB's swallowed exceptions for EnvCell IDs.

## Cause(s)

[Replace this section with the captured findings. Example shape:]

Two distinct failure modes captured at Holtburg:

1. **`KeyNotFoundException` — N cells affected** — Exception thrown from `PrepareCellStructMeshData` line XXX when trying to look up surface `0x08001234`. Affected cells: `0xA9B40100`, `0xA9B40111`, ...

2. **`NULL_RESULT` — M cells affected** — WB's `ResolveId` returned empty for `EnvironmentId 0xD000XXXX`, causing `PrepareEnvCellMeshData` to skip the cellGeometry branch and produce an empty result. Affected cells: ...

[OR if only one cause is observed:]

Single failure mode: [exception type] thrown in [location] for all 26 cells. Root cause: [analysis].

## Sample log lines

```
[paste 5-10 actual captured FAILED / NULL_RESULT lines here]
```

## Proposed fix

[Concrete code change for each distinct cause. For example:]

- For `KeyNotFoundException` on surface lookup: add a null-guard in `WbMeshAdapter.PopulateMetadata` AND skip the failing surface in our acdream-side processing.
- For `NULL_RESULT` from missing `EnvironmentId`: log + skip with a sentinel render-data so the dispatcher gracefully draws nothing instead of failing silently.

Each fix is a single-file change. Task 4 of this plan implements them.

## Verification approach

After Task 4's fix:
- Re-launch with the same probe flag.
- Confirm previously-failing cells now emit `[indoor-upload] completed` lines.
- Visual: floor renders in Holtburg Inn.
```

- [ ] **Step 2: Commit**

```bash
git add docs/research/2026-05-19-indoor-cell-rendering-cause.md
git commit -m "$(cat <<'EOF'
docs(research): Phase 2 cause report — <one-line summary of finding>

Captured at Holtburg with the ContinueWith-based exception surfacer
from Task 1. <Describe finding in 2-3 sentences: which exception types
fired, for how many cells, the root cause.>

Fix shape decided: <one sentence>. Implemented in next commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)"
```

---

## Task 4: Apply the targeted fix

**The fix shape is unknown until Task 2 captures.** This task's code is data-driven. The plan below lists the four most likely fix shapes; the implementer picks the matching one(s) and implements them.

### 4a — If the cause is `KeyNotFoundException` / missing dat record

Most likely path: WB's `PrepareCellStructMeshData` calls `_dats.Portal.TryGet<Surface>(surfaceId, out var surface)`, gets `false`, then crashes when later code assumes non-null.

**Files:**
- Modify: TBD by exception stack — likely a WB fork patch OR a guard at our acdream call site.

- [ ] **Step 1: Open the throwing file based on the exception stack trace**

The probe line will show:
```
stack=[at PrepareCellStructMeshData in ObjectMeshManager.cs:line | at PrepareEnvCellMeshData in ObjectMeshManager.cs:line | ...]
```

Open that file at that line. Confirm the missing-dat-record assumption.

- [ ] **Step 2: Patch shape (WB fork, if in WB)**

In `references/WorldBuilder/Chorizite.OpenGLSDLBackend/Lib/ObjectMeshManager.cs`, add a null-guard at the throwing line:

```csharp
// Pre-Phase-2: WB assumed every surface in envCell.Surfaces was
// resolvable. Some Holtburg cells reference surfaces that aren't in the
// loaded portal dat, causing a NullRef in the throwing line below.
// Guard: skip the surface if it doesn't resolve.
if (!_dats.Portal.TryGet<Surface>(surfaceId, out var surface))
{
    continue;  // or: surface = _fallbackSurface; whichever fits
}
```

(Exact code depends on the stack. The implementer reads the actual throwing line and adapts.)

- [ ] **Step 3: Build, capture, verify**

```bash
dotnet build src/AcDream.App/AcDream.App.csproj -c Debug
```

Then re-run Task 2's launch + capture. Confirm:
- Previously-failing cells now have `[indoor-upload] completed` lines.
- No new `[indoor-upload] FAILED` lines for those cells.

- [ ] **Step 4: Commit**

```bash
git add references/WorldBuilder/Chorizite.OpenGLSDLBackend/Lib/ObjectMeshManager.cs
# OR whatever file was patched
git commit -m "$(cat <<'EOF'
fix(wb): null-guard for missing surface in PrepareCellStructMeshData

Phase 2 capture found <N> Holtburg cells silently failing with
<ExceptionType> thrown at <file>:<line> when WB tried to look up
surface 0x... that isn't resolvable in the loaded portal dat.

Patch: <one-sentence description of the guard>.

Visual-verified: floor now renders in Holtburg Inn.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)"
```

### 4b — If the cause is `NULL_RESULT` (clean null return from WB)

WB's `PrepareMeshData` returns null without throwing. Examined paths in the WB source:
- Line 568: `_dats.Portal.TryGet<Environment>(envId, ...)` fails → returns null.
- Line 583: `type == DBObjType.Unknown` (ResolveId didn't classify the record) → returns null.

**Files:**
- Modify: probably WbMeshAdapter to detect and log, then either accept the cell as "no geometry" gracefully OR investigate the dat issue.

- [ ] **Step 1: Read which path triggered**

Look at the `NULL_RESULT` cells' EnvironmentId values. If the EnvironmentId looks corrupt or out of range, the dat is the issue. If it looks valid, WB's `ResolveId` is broken for that record.

- [ ] **Step 2: Add a guard at our acdream call site OR patch WB**

Depending on the finding:
- **If dat is genuinely missing data**: skip the cell with a warning. Don't try to render its mesh. Log once via memory.
- **If WB's ResolveId mis-classifies**: patch WB or work around by pre-checking with our own `_dats.Get<EnvCell>(envCellId)` before calling `IncrementRefCount`.

- [ ] **Step 3: Build, capture, verify, commit** (same pattern as 4a Step 3-4).

### 4c — If the cause is a `NullReferenceException` in our code path

Less likely but possible — if `PopulateMetadata` or `CellMesh.Build` crashes when invoked from a worker thread.

**Files:**
- Modify: the specific acdream file the stack trace points to.

- [ ] **Step 1: Read the throwing line**
- [ ] **Step 2: Add the appropriate null-guard**
- [ ] **Step 3: Build, capture, verify, commit.**

### 4d — If the cause is something else entirely

If the captured exception type doesn't match 4a-4c, **STOP and re-design**. The fix shape needs the implementer's judgment + possibly a fresh brainstorm session. Don't paper over the cause with a generic try/catch.

---

## Task 5: Verification + visualization

**Files:**
- Create: `docs/research/2026-05-19-indoor-cell-rendering-verification.md`

- [ ] **Step 1: Re-launch with the probe and re-walk Holtburg**

Same as Task 2 Steps 2-4, but expectation flipped: `[indoor-upload] FAILED` / `NULL_RESULT` lines for previously-failing cells should NOT appear; `[indoor-upload] completed` lines should appear instead.

- [ ] **Step 2: Visual verification by user**

User walks into Holtburg Inn AND the other buildings whose cells were previously missing. Expected: floors visible, no missing geometry.

- [ ] **Step 3: Write the verification report**

Create the file documenting:

```markdown
# Indoor Cell Rendering — Phase 2 Verification

**Date:** 2026-05-19
**Outcome:** Floor renders in Holtburg Inn.

## Probe re-capture

After Task 4's fix:
- Previously-failing cells: <list — e.g. `0xA9B40100`, `0xA9B40111`, ...>
- Now emit `[indoor-upload] completed cellId=0x... isSetup=True hasEnvCellGeom=True cellGeomVerts=<N> uploadOk=True`
- No new `[indoor-upload] FAILED` or `NULL_RESULT` lines for these cells.

## Visual confirmation

User walked into:
- Holtburg Inn — floor visible. ✓
- <other buildings tested> — floor visible. ✓

## Regressions checked

- Outdoor terrain still renders correctly. ✓
- NPCs, mobs, scenery still render. ✓
- No new build warnings, no new test failures.

## Closes

This concludes Phase 2 of the indoor cell rendering fix.
```

- [ ] **Step 4: Commit**

```bash
git add docs/research/2026-05-19-indoor-cell-rendering-verification.md
git commit -m "$(cat <<'EOF'
docs(research): Phase 2 verification — floor renders in Holtburg Inn

Post-fix re-capture confirms previously-failing cells now emit
[indoor-upload] completed. Visual verification by user confirms
floors visible in Holtburg Inn and <other tested buildings>.

Phase 2 complete.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)"
```

---

## Task 6: Roadmap update

**Files:**
- Modify: `docs/plans/2026-04-11-roadmap.md`

- [ ] **Step 1: Read the roadmap's "shipped" section**

Open `docs/plans/2026-04-11-roadmap.md`. Find the section listing recently-shipped phases (likely near the top, in a "shipped" table or chronological list).

- [ ] **Step 2: Add an entry for Phase 2 indoor cell rendering fix**

Add an entry matching the existing pattern of shipped-row entries. Example shape:

```markdown
| <next row number> | 2026-05-19 | Indoor cell rendering — Phase 1 (diagnostics) + Phase 2 (fix) | Surfaced + fixed WB's silent failure for 26/123 Holtburg cells. Spec at [phase 1](../superpowers/specs/2026-05-19-indoor-cell-rendering-fix-design.md) + [phase 2](../superpowers/specs/2026-05-19-phase2-indoor-cell-rendering-fix-design.md). Cause: <one-line>. Fix: <one-line>. Visual-verified at Holtburg Inn. |
```

(Read the actual existing row format and match it.)

- [ ] **Step 3: Commit**

```bash
git add docs/plans/2026-04-11-roadmap.md
git commit -m "$(cat <<'EOF'
docs(roadmap): Phase 2 indoor cell rendering fix shipped

Phase 1 diagnostics + Phase 2 fix landed today. Indoor floor rendering
restored for Holtburg cells previously missing due to WB silent
failure. Spec, plan, and verification documents committed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)"
```

---

## Acceptance Criteria

- [ ] Task 1 commits: `WbMeshAdapter.IncrementRefCount` attaches the continuation. `dotnet build` clean.
- [ ] Task 2 capture: `[indoor-upload] FAILED` or `NULL_RESULT` lines fire for previously-failing cells. Distinct cause(s) identified.
- [ ] Task 3 cause report: documented in `docs/research/2026-05-19-indoor-cell-rendering-cause.md`.
- [ ] Task 4 fix: applied + committed. Build clean. Tests clean (no new failures; pre-existing 8 physics/input failures unchanged).
- [ ] Task 5 verification: post-fix probe re-capture confirms `[indoor-upload] completed` for previously-failing cells. User visually confirms floor renders in Holtburg Inn.
- [ ] Task 6 roadmap update: shipped row added.

---

## Subagent dispatch notes

- **Task 1** is mechanical (well-specified code edit) — dispatch to Sonnet.
- **Task 2** is operator-driven — the controller (parent) drives the launch + capture, not a subagent. The user MUST walk the client.
- **Task 3** is analytical (interpret captured data) — controller writes inline, or dispatch a Sonnet subagent with the captured log as context.
- **Task 4** is judgment-intensive (fix shape depends on data) — controller writes inline. If complex, a fresh brainstorm may be needed.
- **Task 5** is similar to Task 2 (user-driven walk + analysis).
- **Task 6** is mechanical — dispatch to Sonnet OR controller writes inline.