[Turbopack] Explore opt-in shared server output for co-deployed context-neutral graphs

Link to the code that reproduces this issue

https://github.com/Ustay-coder/next-turbopack-cross-context-sharing-repro/tree/b4df11ee2150e8d23a997a27ac989975c1f59c39

To Reproduce

1. Clone the public minimal App Router reproduction and check out the pinned commit b4df11ee2150e8d23a997a27ac989975c1f59c39.
2. Install dependencies with pnpm 9.0.0.
3. Build with next@16.3.1-canary.0 on Linux x64 (Turbopack is the default bundler).
4. The app imports the same deterministic, context-neutral ESM/JSON graph from both an app-route entry and an app-rsc entry.
5. Run the four-cell factorial measurement (baseline / route / action / both, three runs) documented in results/linux-x64.md and compare the gzip sizes of the emitted .next/server output.

Both runtime paths pass. The full methodology and results are in the Linux report.

Current vs. Expected behavior

Current: With Turbopack, the second context's server output retains a full second copy of the shared graph. Measured second-context gzip cost: 181,504 B (duplication ratio 0.999994). With Webpack, the same graph adds -1,043 B (effectively zero duplication). Both runtime paths pass, so this is a retained-output-size observation, not a claim about duplicate execution within one runtime.

Expected: This is deliberately an Ideas proposal, not a bug report. The question is whether Turbopack could offer an opt-in server-output strategy for a context-neutral graph when a deployment target can prove compatible runtime semantics and co-location, while preserving self-contained server endpoint output as the default. The output-size versus independent-deployment trade-off should be explicit for server contexts, rather than leaving applications with client-only chunking controls for this boundary.

Provide environment information

From the reproduction's Linux report (`results/linux-x64.md`):

Operating System:
  Platform: linux
  Arch: x64

Binaries:
  Node: v22.23.2
  pnpm: 9.0.0

Relevant Packages:
  next: 16.3.1-canary.0 (tag commit 5005bd083874d366f95fd34da7a5d27837cbd5fa)
  react: 19.2.8

Which area(s) are affected? (Select all that apply)

Turbopack, Output

Which stage(s) are affected? (Select all that apply)

next build (local)

Additional context

Goals

1. Preserve self-contained server endpoint output as the default, so routes can continue to be deployed independently when a platform requires that.
2. Clarify whether Turbopack could offer an opt-in server-output strategy for a context-neutral graph when a deployment target can prove compatible runtime semantics and co-location.
3. Make the output-size versus independent-deployment trade-off explicit for server contexts, rather than leaving applications with client-only chunking controls for this boundary.

Non-Goals

1. This is not a request to merge app-route and app-rsc into one compilation context, or to change the default Lambda/function packaging model.
2. This is not a request to share modules across different runtime targets, resolution conditions, transforms, externalization choices, side-effect semantics, or availability constraints.
3. This is not the instanceof / singleton correctness issue tracked in #89192, and it does not require client chunking, HMR, Pages Router, or Edge/Node cross-runtime work.

Background

#89192 establishes an important constraint: multiple physical server outputs can be expected because different routes may be deployed as separate production Lambdas. I agree that self-contained endpoint output should remain the default.

The remaining question is about an optional output-size trade-off when that deployment constraint does not apply. I prepared a public, minimal App Router reproduction pinned to an immutable commit:

https://github.com/Ustay-coder/next-turbopack-cross-context-sharing-repro/tree/b4df11ee2150e8d23a997a27ac989975c1f59c39

It imports the same deterministic, context-neutral ESM/JSON graph from both an app-route entry and an app-rsc entry. On Linux x64 with next@16.3.1-canary.0, a four-cell factorial measurement with three runs found:

Bundler	Second-context gzip cost	Duplication ratio
Turbopack	181,504 B	0.999994
Webpack	-1,043 B	effectively zero

The full methodology and results are in the Linux report. Both runtime paths pass. This is therefore a retained-output-size observation, not a claim about duplicate execution within one runtime.

I also checked the recently added experimental.turbopackChunking controls. They expose client chunk-group merging choices, but I could not find an equivalent server-output strategy for this app-route / app-rsc boundary.

Proposal

Would the team be open to defining an opt-in mechanism for this trade-off, while keeping independent, self-contained server outputs as the default?

I am intentionally not proposing a source-hash-only deduplication algorithm. Any sharing would need to establish equivalence of at least the emitted module factory, resolution/condition set, transforms, dependency edges, runtime, externalization, side effects, and async availability.

Two possible directions seem worth discussing:

1. Adapter-consumable manifest or shared artifact. Next.js could describe a reusable server graph and its requirements, while a deployment adapter that guarantees co-location could choose to package it once. Without such an adapter, endpoint outputs would remain self-contained.
2. Explicit experimental server-output strategy. Turbopack could expose a narrowly scoped, deployment-aware mode for compatible Node server contexts. This would make the size trade-off visible to users, but would carry a broader support and compatibility surface.

Would either direction fit the intended Next.js boundary, or is this fundamentally a deployment-adapter concern that should be documented outside Turbopack? If the team sees a viable contract, I can contribute a narrowly scoped regression fixture after the supported contexts and semantic exclusions are agreed.

---

Start by running the pinned reproduction and reading results/linux-x64.md, then review the experimental.turbopackChunking controls in PR #96398. Compare the reported Turbopack and Webpack output behavior and determine whether a supported server-output contract or deployment-adapter boundary can be defined; done requires an agreed direction before implementing a regression fixture.