We're converging well. Let me confirm agreements, raise final concerns, and request the definitive merged architecture.

## CONFIRMED AGREEMENTS

1. **Reference ordering:** Scene (weight 1-3) → Pose/Flash (weight 4-5) → Expression Transfer (weight 6-7) → Character Identity (weight 8-10) → Prompt Text. Recency bias = identity dominance.

2. **Native 9:16 production:** Triptychs demoted to optional QC tool. All production frames generated natively at delivery aspect ratio.

3. **3-Pass Pipeline:** ENV anchor (3-pro) → Flash exploration x4 (3.1-flash) → Pro final render (3-pro). Flash candidate serves as lighting/pose anchor.

4. **Reference mirroring:** Dynamic `ImageOps.mirror()` for screen direction enforcement. Hardware fix > text prompt.

5. **Expression Transfer:** Use stock/generated emotion reference images labeled `[FACIAL EXPRESSION TO MATCH]` at medium weight, keeping Jinx identity refs pristine at high weight.

6. **Color contamination mitigation:** White-bg identity refs stay pristine. Flash exploration output (correctly lit via scene ref) serves as the lighting anchor for the final Pro render.

7. **Wide-shot prompt branching:** Remove facial detail demands from WIDE/LS prompts. Accept impressionistic faces at distance. FaceDetailer as optional post-process.

8. **Positive constraints:** "Anatomically flawless hands, exactly five fingers" > "no deformed hands."

9. **Kinetic descriptors:** "Motion blur, off-axis framing, dust kicked into lens" > "CHARACTER IS PUSHING HARD."

10. **Lighting vector locking:** Explicit directional coordinates ("amber light hard shadows TOP LEFT") in every prompt within a scene.

## REMAINING CONCERNS FOR FINAL ROUND

### 1. Expression Reference Library — Bootstrap Problem
Expression Transfer requires stock/generated emotion reference images. Where do these come from at Day 1? Options:
a) Pre-generate a library of ~20 expressions using Flash (cost: ~$0.78) with generic actors
b) Source from stock photography (license concerns)
c) Generate on-demand per shot (adds latency)

I lean toward (a) — pre-generate a canonical expression library indexed by emotion keyword (terror, exhaustion, rage, grief, wonder, determination, etc.) and store in `starsend/assets/expressions/`. One-time cost.

### 2. Two-Character Shots — Reference Budget
EP001 has dual-character shots (Jinx + Kian) in the later beats. Your API docs say 6 object + 5 character references max. For a two-character shot with Expression Transfer:
- Scene ref: 1
- Flash pose ref: 1
- Expression ref per character: 2
- Identity refs per character: 2-3 each (4-6 total)
- Total: 8-10 references

Is this within budget? If not, which refs get cut first?

### 3. Cost Optimization for Tier 1 (Pipeline) Shots
The production sprint classifies ~60-70% of shots as Tier 1 (pipeline-handled). For these "standard" shots (e.g., MS of Jinx working, CU of a prop), do we always need the Flash exploration pass? Or can we skip straight to Pro for simple compositions where the storyboard is highly prescriptive?

Proposed: Add a `complexity` field to the shot pipeline config:
- `simple` → Skip Flash exploration, go straight to Pro ($0.134/shot)
- `standard` → Full 3-pass pipeline (~$0.30/shot)
- `complex` → 3-pass + Expression Transfer + multiple Pro attempts (~$0.50/shot)

### 4. Batch Efficiency — Scene-Level Amortization
Within a scene (e.g., shots 1-10 all in the same corridor), the ENV anchor only runs once. But does the Flash exploration improve if we chain: Feed the selected Flash hero from Shot N as an additional scene ref for Shot N+1? This would create a progressive consistency chain similar to the original scene reference chaining, but at the Flash exploration level.

### 5. gemini-3.1-flash-image-preview Availability
You recommended 3.1 Flash for exploration. I confirmed it's in the model list (`gemini-3.1-flash-image-preview`). Is there a pricing difference from 2.5 Flash? Any known limitations (max resolution, reference count)?

## REQUESTS FOR FINAL ROUND

1. Definitive merged architecture summary with file-by-file responsibilities
2. Top 5 implementation risks and mitigations
3. Specific test protocol for EP001 shots 1-3 (the same shots we tested in V4)
4. Cost model for EP001 (31 shots) under the merged pipeline
5. Any final "undocumented tricks" you haven't mentioned yet