Worker-side truncation is invisible to the seam's `truncated` flag: a capped non-streaming exec returns a real exit code over cut output

[pdettori]

Problem

PR #188 made truncation explicit on the SandboxTransport seam (truncated: boolean, with the invariant truncated === true ⇒ exitCode === null). One path can still cut output and report success: the Go worker's own buffer cap.

remote-worker/internal/exec/runner.go:295-296, on the non-streaming path (runner.go:233, if !s.Streaming):

} else if room := BufferCap - buf.Len(); room > 0 {
    buf.Write(b[:min(n, room)])
}

Output past BufferCap is dropped silently, and the worker then sends End carrying the command's real exit code. GrpcRelayTransport receives exactly BufferCap bytes — which is not > outputCap, since BufferCap == DEFAULT_OUTPUT_CAP == 8 MiB — so it resolves { truncated: false, exitCode: <real code> } over output the worker truncated.

The equality of the two caps is precisely what makes it silent. If they differed by a byte in either direction, the harness would either trip its own cap or never fill the buffer.

Reachability

Not reachable from this repo today. packages/k8s-sandbox/src/grpc-relay-transport.ts:51 hardcodes streaming: true on every Exec, so the non-streaming branch is never taken by our client.

But runner.go:36 warns about exactly this: Exec.streaming is relay-supplied and false is the proto3 default, so any other relay or client — buggy, third-party, or hostile — reaches this path with no privilege at all. sandbox/v1 is a language-neutral contract intended for third-party workers, so "our client happens not to do it" is a weak guarantee.

Why it matters

This is the one remaining path where the seam's contract is violated rather than merely qualified: a caller gets a success exit code over incomplete output, with no signal. That is the same defect class as #181 (a killed bash reported as success), which #188 just closed on the other transports.

Options

1. Make the worker signal truncation on the wire. Requires a field on End (or a distinguished exit code) and a sandbox/v1 change, so it is a contract decision across three languages — the same class of change req_id is not globally unique: colliding ids from multiple harness replicas misroute frames at the relay #179 weighed.
2. Make the caps deliberately unequal, so the harness's cap always trips first (e.g. worker BufferCap = harness cap + 1 chunk). Cheap and no proto change, but it encodes a subtle invariant in two repos that packages/k8s-sandbox/test/output-cap-coupling.test.ts currently pins as equal — that test would need to pin the offset instead.
3. Reject streaming: false at the relay, since our client never sets it. Narrows the surface without fixing the contract for third-party workers.

Refs

Found by the whole-branch review of #188 (ST6, docs/specs/2026-08-30-seam-output-cap-truncation-design.md), whose §2 explicitly excluded worker changes. Pre-existing, not introduced by that branch. Related: #181, epic #89.